KenLM df5be22 lmplz for estimation

42 files changed, 4031 insertions, 56 deletions

diff --git a/lm/Jamfile b/lm/Jamfile
index a4ec75002..0faea30f4 100644
--- a/lm/Jamfile
+++ b/lm/Jamfile
@@ -13,7 +13,7 @@ update-if-changed $(ORDER-LOG) $(max-order) ;
 
 max-order += <dependency>$(ORDER-LOG) ;
 
-fakelib kenlm : bhiksha.cc binary_format.cc config.cc lm_exception.cc model.cc quantize.cc read_arpa.cc search_hashed.cc search_trie.cc trie.cc trie_sort.cc value_build.cc virtual_interface.cc vocab.cc ../util//kenutil : <include>.. $(max-order) : : <include>.. $(max-order) ;
+fakelib kenlm : bhiksha.cc binary_format.cc config.cc lm_exception.cc model.cc quantize.cc read_arpa.cc search_hashed.cc search_trie.cc sizes.cc trie.cc trie_sort.cc value_build.cc virtual_interface.cc vocab.cc ../util//kenutil : <include>.. $(max-order) : : <include>.. $(max-order) ;
 
 import testing ;
 
@@ -26,4 +26,4 @@ exe build_binary : build_binary.cc kenlm ../util//kenutil ;
 exe kenlm_max_order : max_order.cc : <include>.. $(max-order) ;
 exe fragment : fragment.cc kenlm ;
 
-alias programs : query build_binary kenlm_max_order fragment ;
+alias programs : query build_binary kenlm_max_order fragment filter//filter : <threading>multi:<source>builder//lmplz ;
diff --git a/lm/build_binary.cc b/lm/build_binary.cc
index 2b8c9d5b2..ab2c0c320 100644
--- a/lm/build_binary.cc
+++ b/lm/build_binary.cc
@@ -1,10 +1,14 @@
 #include "lm/model.hh"
+#include "lm/sizes.hh"
 #include "util/file_piece.hh"
+#include "util/usage.hh"
 
+#include <algorithm>
 #include <cstdlib>
 #include <exception>
 #include <iostream>
 #include <iomanip>
+#include <limits>
 
 #include <math.h>
 #include <stdlib.h>
@@ -19,8 +23,8 @@ namespace lm {
 namespace ngram {
 namespace {
 
-void Usage(const char *name) {
-  std::cerr << "Usage: " << name << " [-u log10_unknown_probability] [-s] [-i] [-w mmap|after] [-p probing_multiplier] [-t trie_temporary] [-m trie_building_megabytes] [-q bits] [-b bits] [-a bits] [type] input.arpa [output.mmap]\n\n"
+void Usage(const char *name, const char *default_mem) {
+  std::cerr << "Usage: " << name << " [-u log10_unknown_probability] [-s] [-i] [-w mmap|after] [-p probing_multiplier] [-T trie_temporary] [-S trie_building_mem] [-q bits] [-b bits] [-a bits] [type] input.arpa [output.mmap]\n\n"
 "-u sets the log10 probability for <unk> if the ARPA file does not have one.\n"
 "   Default is -100.  The ARPA file will always take precedence.\n"
 "-s allows models to be built even if they do not have <s> and </s>.\n"
@@ -38,8 +42,11 @@ void Usage(const char *name) {
 "trie is a straightforward trie with bit-level packing.  It uses the least\n"
 "memory and is still faster than SRI or IRST.  Building the trie format uses an\n"
 "on-disk sort to save memory.\n"
-"-t is the temporary directory prefix.  Default is the output file name.\n"
-"-m limits memory use for sorting.  Measured in MB.  Default is 1024MB.\n"
+"-T is the temporary directory prefix.  Default is the output file name.\n"
+"-S determines memory use for sorting.  Default is " << default_mem << ".  This is compatible\n"
+"   with GNU sort.  The number is followed by a unit: \% for percent of physical\n"
+"   memory, b for bytes, K for Kilobytes, M for megabytes, then G,T,P,E,Z,Y.  \n"
+"   Default unit is K for Kilobytes.\n"
 "-q turns quantization on and sets the number of bits (e.g. -q 8).\n"
 "-b sets backoff quantization bits.  Requires -q and defaults to that value.\n"
 "-a compresses pointers using an array of offsets.  The parameter is the\n"
@@ -83,47 +90,6 @@ void ParseFileList(const char *from, std::vector<std::string> &to) {
   }
 }
 
-void ShowSizes(const char *file, const lm::ngram::Config &config) {
-  std::vector<uint64_t> counts;
-  util::FilePiece f(file);
-  lm::ReadARPACounts(f, counts);
-  uint64_t sizes[6];
-  sizes[0] = ProbingModel::Size(counts, config);
-  sizes[1] = RestProbingModel::Size(counts, config);
-  sizes[2] = TrieModel::Size(counts, config);
-  sizes[3] = QuantTrieModel::Size(counts, config);
-  sizes[4] = ArrayTrieModel::Size(counts, config);
-  sizes[5] = QuantArrayTrieModel::Size(counts, config);
-  uint64_t max_length = *std::max_element(sizes, sizes + sizeof(sizes) / sizeof(uint64_t));
-  uint64_t min_length = *std::min_element(sizes, sizes + sizeof(sizes) / sizeof(uint64_t));
-  uint64_t divide;
-  char prefix;
-  if (min_length < (1 << 10) * 10) {
-    prefix = ' ';
-    divide = 1;
-  } else if (min_length < (1 << 20) * 10) {
-    prefix = 'k';
-    divide = 1 << 10;
-  } else if (min_length < (1ULL << 30) * 10) {
-    prefix = 'M';
-    divide = 1 << 20;
-  } else {
-    prefix = 'G';
-    divide = 1 << 30;
-  }
-  long int length = std::max<long int>(2, static_cast<long int>(ceil(log10((double) max_length / divide))));
-  std::cout << "Memory estimate:\ntype    ";
-  // right align bytes.  
-  for (long int i = 0; i < length - 2; ++i) std::cout << ' ';
-  std::cout << prefix << "B\n"
-    "probing " << std::setw(length) << (sizes[0] / divide) << " assuming -p " << config.probing_multiplier << "\n"
-    "probing " << std::setw(length) << (sizes[1] / divide) << " assuming -r models -p " << config.probing_multiplier << "\n"
-    "trie    " << std::setw(length) << (sizes[2] / divide) << " without quantization\n"
-    "trie    " << std::setw(length) << (sizes[3] / divide) << " assuming -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits << " quantization \n"
-    "trie    " << std::setw(length) << (sizes[4] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " array pointer compression\n"
-    "trie    " << std::setw(length) << (sizes[5] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits<< " array pointer compression and quantization\n";
-}
-
 void ProbingQuantizationUnsupported() {
   std::cerr << "Quantization is only implemented in the trie data structure." << std::endl;
   exit(1);
@@ -136,11 +102,14 @@ void ProbingQuantizationUnsupported() {
 int main(int argc, char *argv[]) {
   using namespace lm::ngram;
 
+  const char *default_mem = util::GuessPhysicalMemory() ? "80%" : "1G";
+
   try {
     bool quantize = false, set_backoff_bits = false, bhiksha = false, set_write_method = false, rest = false;
     lm::ngram::Config config;
+    config.building_memory = util::ParseSize(default_mem);
     int opt;
-    while ((opt = getopt(argc, argv, "q:b:a:u:p:t:m:w:sir:")) != -1) {
+    while ((opt = getopt(argc, argv, "q:b:a:u:p:t:T:m:S:w:sir:")) != -1) {
       switch(opt) {
         case 'q':
           config.prob_bits = ParseBitCount(optarg);
@@ -161,12 +130,16 @@ int main(int argc, char *argv[]) {
         case 'p':
           config.probing_multiplier = ParseFloat(optarg);
           break;
-        case 't':
+        case 't': // legacy
+        case 'T':
           config.temporary_directory_prefix = optarg;
           break;
-        case 'm':
+        case 'm': // legacy
           config.building_memory = ParseUInt(optarg) * 1048576;
           break;
+        case 'S':
+          config.building_memory = std::min(static_cast<uint64_t>(std::numeric_limits<std::size_t>::max()), util::ParseSize(optarg));
+          break;
         case 'w':
           set_write_method = true;
           if (!strcmp(optarg, "mmap")) {
@@ -174,7 +147,7 @@ int main(int argc, char *argv[]) {
           } else if (!strcmp(optarg, "after")) {
             config.write_method = Config::WRITE_AFTER;
           } else {
-            Usage(argv[0]);
+            Usage(argv[0], default_mem);
           }
           break;
         case 's':
@@ -189,7 +162,7 @@ int main(int argc, char *argv[]) {
           config.rest_function = Config::REST_LOWER;
           break;
         default:
-          Usage(argv[0]);
+          Usage(argv[0], default_mem);
       }
     }
     if (!quantize && set_backoff_bits) {
@@ -212,7 +185,7 @@ int main(int argc, char *argv[]) {
       from_file = argv[optind + 1];
       config.write_mmap = argv[optind + 2];
     } else {
-      Usage(argv[0]);
+      Usage(argv[0], default_mem);
     }
     if (!strcmp(model_type, "probing")) {
       if (!set_write_method) config.write_method = Config::WRITE_AFTER;
@@ -242,7 +215,7 @@ int main(int argc, char *argv[]) {
         }
       }
     } else {
-      Usage(argv[0]);
+      Usage(argv[0], default_mem);
     }
   }
   catch (const std::exception &e) {
diff --git a/lm/builder/Jamfile b/lm/builder/Jamfile
new file mode 100644
index 000000000..9b54911a0
--- /dev/null
+++ b/lm/builder/Jamfile
@@ -0,0 +1,9 @@
+fakelib builder : corpus_count.cc adjust_counts.cc initial_probabilities.cc interpolate.cc pipeline.cc print.cc 
+  ../../util//kenutil ../../util/stream//stream ../../util/double-conversion//double-conversion ..//kenlm
+  : : : <library>/top//boost_thread $(timer-link) ;
+
+exe lmplz : main.cc builder /top//boost_program_options ;
+
+import testing ;
+unit-test corpus_count_test : corpus_count_test.cc builder /top//boost_unit_test_framework ;
+unit-test adjust_counts_test : adjust_counts_test.cc builder /top//boost_unit_test_framework ;
diff --git a/lm/builder/README.md b/lm/builder/README.md
new file mode 100644
index 000000000..be0d35e26
--- /dev/null
+++ b/lm/builder/README.md
@@ -0,0 +1,47 @@
+Dependencies
+============
+
+Boost >= 1.42.0 is required.  
+
+For Ubuntu,
+```bash
+sudo apt-get install libboost1.48-all-dev
+```
+
+Alternatively, you can download, compile, and install it yourself:
+
+```bash
+wget http://sourceforge.net/projects/boost/files/boost/1.52.0/boost_1_52_0.tar.gz/download -O boost_1_52_0.tar.gz
+tar -xvzf boost_1_52_0.tar.gz
+cd boost_1_52_0
+./bootstrap.sh
+./b2
+sudo ./b2 install
+```
+
+Local install options (in a user-space prefix directory) are also possible. See http://www.boost.org/doc/libs/1_52_0/doc/html/bbv2/installation.html.
+
+
+Building
+========
+
+```bash
+bjam
+```
+Your distribution might package bjam and boost-build separately from Boost.  Both are required.   
+
+Usage
+=====
+
+Run
+```bash
+$ bin/lmplz
+```
+to see command line arguments
+
+Running
+=======
+
+```bash
+bin/lmplz -o 5 <text >text.arpa
+```
diff --git a/lm/builder/TODO b/lm/builder/TODO
new file mode 100644
index 000000000..cb5aef3a0
--- /dev/null
+++ b/lm/builder/TODO
@@ -0,0 +1,5 @@
+More tests!
+Sharding.
+Some way to manage all the crazy config options.
+Option to build the binary file directly.  
+Interpolation of different orders.  
diff --git a/lm/builder/adjust_counts.cc b/lm/builder/adjust_counts.cc
new file mode 100644
index 000000000..a6f48011f
--- /dev/null
+++ b/lm/builder/adjust_counts.cc
@@ -0,0 +1,216 @@
+#include "lm/builder/adjust_counts.hh"
+#include "lm/builder/multi_stream.hh"
+#include "util/stream/timer.hh"
+
+#include <algorithm>
+
+namespace lm { namespace builder {
+
+BadDiscountException::BadDiscountException() throw() {}
+BadDiscountException::~BadDiscountException() throw() {}
+
+namespace {
+// Return last word in full that is different.  
+const WordIndex* FindDifference(const NGram &full, const NGram &lower_last) {
+  const WordIndex *cur_word = full.end() - 1;
+  const WordIndex *pre_word = lower_last.end() - 1;
+  // Find last difference.  
+  for (; pre_word >= lower_last.begin() && *pre_word == *cur_word; --cur_word, --pre_word) {}
+  return cur_word;
+}
+
+class StatCollector {
+  public:
+    StatCollector(std::size_t order, std::vector<uint64_t> &counts, std::vector<Discount> &discounts) 
+      : orders_(order), full_(orders_.back()), counts_(counts), discounts_(discounts) {
+      memset(&orders_[0], 0, sizeof(OrderStat) * order);
+    }
+
+    ~StatCollector() {}
+
+    void CalculateDiscounts() {
+      counts_.resize(orders_.size());
+      discounts_.resize(orders_.size());
+      for (std::size_t i = 0; i < orders_.size(); ++i) {
+        const OrderStat &s = orders_[i];
+        counts_[i] = s.count;
+
+        for (unsigned j = 1; j < 4; ++j) {
+          // TODO: Specialize error message for j == 3, meaning 3+
+          UTIL_THROW_IF(s.n[j] == 0, BadDiscountException, "Could not calculate Kneser-Ney discounts for "
+              << (i+1) << "-grams with adjusted count " << (j+1) << " because we didn't observe any "
+              << (i+1) << "-grams with adjusted count " << j << "; Is this small or artificial data?");
+        }
+
+        // See equation (26) in Chen and Goodman.
+        discounts_[i].amount[0] = 0.0;
+        float y = static_cast<float>(s.n[1]) / static_cast<float>(s.n[1] + 2.0 * s.n[2]);
+        for (unsigned j = 1; j < 4; ++j) {
+          discounts_[i].amount[j] = static_cast<float>(j) - static_cast<float>(j + 1) * y * static_cast<float>(s.n[j+1]) / static_cast<float>(s.n[j]);
+          UTIL_THROW_IF(discounts_[i].amount[j] < 0.0 || discounts_[i].amount[j] > j, BadDiscountException, "ERROR: " << (i+1) << "-gram discount out of range for adjusted count " << j << ": " << discounts_[i].amount[j]);
+        }
+      }
+    }
+
+    void Add(std::size_t order_minus_1, uint64_t count) {
+      OrderStat &stat = orders_[order_minus_1];
+      ++stat.count;
+      if (count < 5) ++stat.n[count];
+    }
+
+    void AddFull(uint64_t count) {
+      ++full_.count;
+      if (count < 5) ++full_.n[count];
+    }
+
+  private:
+    struct OrderStat {
+      // n_1 in equation 26 of Chen and Goodman etc
+      uint64_t n[5];
+      uint64_t count;
+    };
+
+    std::vector<OrderStat> orders_;
+    OrderStat &full_;
+
+    std::vector<uint64_t> &counts_;
+    std::vector<Discount> &discounts_;
+};
+
+// Reads all entries in order like NGramStream does.  
+// But deletes any entries that have <s> in the 1st (not 0th) position on the
+// way out by putting other entries in their place.  This disrupts the sort
+// order but we don't care because the data is going to be sorted again.  
+class CollapseStream {
+  public:
+    CollapseStream(const util::stream::ChainPosition &position) :
+      current_(NULL, NGram::OrderFromSize(position.GetChain().EntrySize())),
+      block_(position) {
+      StartBlock();
+    }
+
+    const NGram &operator*() const { return current_; }
+    const NGram *operator->() const { return &current_; }
+
+    operator bool() const { return block_; }
+
+    CollapseStream &operator++() {
+      assert(block_);
+      if (current_.begin()[1] == kBOS && current_.Base() < copy_from_) {
+        memcpy(current_.Base(), copy_from_, current_.TotalSize());
+        UpdateCopyFrom();
+      }
+      current_.NextInMemory();
+      uint8_t *block_base = static_cast<uint8_t*>(block_->Get());
+      if (current_.Base() == block_base + block_->ValidSize()) {
+        block_->SetValidSize(copy_from_ + current_.TotalSize() - block_base);
+        ++block_;
+        StartBlock();
+      }
+      return *this;
+    }
+
+  private:
+    void StartBlock() {
+      for (; ; ++block_) {
+        if (!block_) return;
+        if (block_->ValidSize()) break;
+      }
+      current_.ReBase(block_->Get());
+      copy_from_ = static_cast<uint8_t*>(block_->Get()) + block_->ValidSize();
+      UpdateCopyFrom();
+    }
+
+    // Find last without bos.  
+    void UpdateCopyFrom() {
+      for (copy_from_ -= current_.TotalSize(); copy_from_ >= current_.Base(); copy_from_ -= current_.TotalSize()) {
+        if (NGram(copy_from_, current_.Order()).begin()[1] != kBOS) break;
+      }
+    }
+
+    NGram current_;
+
+    // Goes backwards in the block
+    uint8_t *copy_from_;
+
+    util::stream::Link block_;
+};
+
+} // namespace
+
+void AdjustCounts::Run(const ChainPositions &positions) {
+  UTIL_TIMER("(%w s) Adjusted counts\n");
+
+  const std::size_t order = positions.size();
+  StatCollector stats(order, counts_, discounts_);
+  if (order == 1) {
+    // Only unigrams.  Just collect stats.  
+    for (NGramStream full(positions[0]); full; ++full) 
+      stats.AddFull(full->Count());
+    stats.CalculateDiscounts();
+    return;
+  }
+
+  NGramStreams streams;
+  streams.Init(positions, positions.size() - 1);
+  CollapseStream full(positions[positions.size() - 1]);
+
+  // Initialization: <unk> has count 0 and so does <s>.  
+  NGramStream *lower_valid = streams.begin();
+  streams[0]->Count() = 0;
+  *streams[0]->begin() = kUNK;
+  stats.Add(0, 0);
+  (++streams[0])->Count() = 0;
+  *streams[0]->begin() = kBOS;
+  // not in stats because it will get put in later.  
+
+  // iterate over full (the stream of the highest order ngrams)
+  for (; full; ++full) {
+    const WordIndex *different = FindDifference(*full, **lower_valid);
+    std::size_t same = full->end() - 1 - different;
+    // Increment the adjusted count.  
+    if (same) ++streams[same - 1]->Count();
+
+    // Output all the valid ones that changed.  
+    for (; lower_valid >= &streams[same]; --lower_valid) {
+      stats.Add(lower_valid - streams.begin(), (*lower_valid)->Count());
+      ++*lower_valid;
+    }
+
+    // This is here because bos is also const WordIndex *, so copy gets
+    // consistent argument types.  
+    const WordIndex *full_end = full->end();
+    // Initialize and mark as valid up to bos.  
+    const WordIndex *bos;
+    for (bos = different; (bos > full->begin()) && (*bos != kBOS); --bos) {
+      ++lower_valid;
+      std::copy(bos, full_end, (*lower_valid)->begin());
+      (*lower_valid)->Count() = 1;
+    }
+    // Now bos indicates where <s> is or is the 0th word of full.  
+    if (bos != full->begin()) {
+      // There is an <s> beyond the 0th word.  
+      NGramStream &to = *++lower_valid;
+      std::copy(bos, full_end, to->begin());
+      to->Count() = full->Count();  
+    } else {
+      stats.AddFull(full->Count());
+    }
+    assert(lower_valid >= &streams[0]);
+  }
+
+  // Output everything valid.
+  for (NGramStream *s = streams.begin(); s <= lower_valid; ++s) {
+    stats.Add(s - streams.begin(), (*s)->Count());
+    ++*s;
+  }
+  // Poison everyone!  Except the N-grams which were already poisoned by the input.   
+  for (NGramStream *s = streams.begin(); s != streams.end(); ++s)
+    s->Poison();
+
+  stats.CalculateDiscounts();
+
+  // NOTE: See special early-return case for unigrams near the top of this function
+}
+
+}} // namespaces
diff --git a/lm/builder/adjust_counts.hh b/lm/builder/adjust_counts.hh
new file mode 100644
index 000000000..f38ff79db
--- /dev/null
+++ b/lm/builder/adjust_counts.hh
@@ -0,0 +1,44 @@
+#ifndef LM_BUILDER_ADJUST_COUNTS__
+#define LM_BUILDER_ADJUST_COUNTS__
+
+#include "lm/builder/discount.hh"
+#include "util/exception.hh"
+
+#include <vector>
+
+#include <stdint.h>
+
+namespace lm {
+namespace builder {
+
+class ChainPositions;
+
+class BadDiscountException : public util::Exception {
+  public:
+    BadDiscountException() throw();
+    ~BadDiscountException() throw();
+};
+
+/* Compute adjusted counts.  
+ * Input: unique suffix sorted N-grams (and just the N-grams) with raw counts.
+ * Output: [1,N]-grams with adjusted counts.  
+ * [1,N)-grams are in suffix order
+ * N-grams are in undefined order (they're going to be sorted anyway).
+ */
+class AdjustCounts {
+  public:
+    AdjustCounts(std::vector<uint64_t> &counts, std::vector<Discount> &discounts)
+      : counts_(counts), discounts_(discounts) {}
+
+    void Run(const ChainPositions &positions);
+
+  private:
+    std::vector<uint64_t> &counts_;
+    std::vector<Discount> &discounts_;
+};
+
+} // namespace builder
+} // namespace lm
+
+#endif // LM_BUILDER_ADJUST_COUNTS__
+
diff --git a/lm/builder/adjust_counts_test.cc b/lm/builder/adjust_counts_test.cc
new file mode 100644
index 000000000..68b5f33ea
--- /dev/null
+++ b/lm/builder/adjust_counts_test.cc
@@ -0,0 +1,106 @@
+#include "lm/builder/adjust_counts.hh"
+
+#include "lm/builder/multi_stream.hh"
+#include "util/scoped.hh"
+
+#include <boost/thread/thread.hpp>
+#define BOOST_TEST_MODULE AdjustCounts
+#include <boost/test/unit_test.hpp>
+
+namespace lm { namespace builder { namespace {
+
+class KeepCopy {
+  public:
+    KeepCopy() : size_(0) {}
+
+    void Run(const util::stream::ChainPosition &position) {
+      for (util::stream::Link link(position); link; ++link) {
+        mem_.call_realloc(size_ + link->ValidSize());
+        memcpy(static_cast<uint8_t*>(mem_.get()) + size_, link->Get(), link->ValidSize());
+        size_ += link->ValidSize();
+      }
+    }
+
+    uint8_t *Get() { return static_cast<uint8_t*>(mem_.get()); }
+    std::size_t Size() const { return size_; }
+
+  private:
+    util::scoped_malloc mem_;
+    std::size_t size_;
+};
+
+struct Gram4 {
+  WordIndex ids[4];
+  uint64_t count;
+};
+
+class WriteInput {
+  public:
+    void Run(const util::stream::ChainPosition &position) {
+      NGramStream input(position);
+      Gram4 grams[] = {
+        {{0,0,0,0},10},
+        {{0,0,3,0},3},
+        // bos
+        {{1,1,1,2},5},
+        {{0,0,3,2},5},
+      };
+      for (size_t i = 0; i < sizeof(grams) / sizeof(Gram4); ++i, ++input) {
+        memcpy(input->begin(), grams[i].ids, sizeof(WordIndex) * 4);
+        input->Count() = grams[i].count;
+      }
+      input.Poison();
+    }
+};
+
+BOOST_AUTO_TEST_CASE(Simple) {
+  KeepCopy outputs[4];
+  std::vector<uint64_t> counts;
+  std::vector<Discount> discount;
+  {
+    util::stream::ChainConfig config;
+    config.total_memory = 100;
+    config.block_count = 1;
+    Chains chains(4);
+    for (unsigned i = 0; i < 4; ++i) {
+      config.entry_size = NGram::TotalSize(i + 1);
+      chains.push_back(config);
+    }
+
+    chains[3] >> WriteInput();
+    ChainPositions for_adjust(chains);
+    for (unsigned i = 0; i < 4; ++i) {
+      chains[i] >> boost::ref(outputs[i]);
+    }
+    chains >> util::stream::kRecycle;
+    BOOST_CHECK_THROW(AdjustCounts(counts, discount).Run(for_adjust), BadDiscountException);
+  }
+  BOOST_REQUIRE_EQUAL(4UL, counts.size());
+  BOOST_CHECK_EQUAL(4UL, counts[0]);
+  // These are no longer set because the discounts are bad.  
+/*  BOOST_CHECK_EQUAL(4UL, counts[1]);
+  BOOST_CHECK_EQUAL(3UL, counts[2]);
+  BOOST_CHECK_EQUAL(3UL, counts[3]);*/
+  BOOST_REQUIRE_EQUAL(NGram::TotalSize(1) * 4, outputs[0].Size());
+  NGram uni(outputs[0].Get(), 1);
+  BOOST_CHECK_EQUAL(kUNK, *uni.begin());
+  BOOST_CHECK_EQUAL(0ULL, uni.Count());
+  uni.NextInMemory();
+  BOOST_CHECK_EQUAL(kBOS, *uni.begin());
+  BOOST_CHECK_EQUAL(0ULL, uni.Count());
+  uni.NextInMemory();
+  BOOST_CHECK_EQUAL(0UL, *uni.begin());
+  BOOST_CHECK_EQUAL(2ULL, uni.Count());
+  uni.NextInMemory();
+  BOOST_CHECK_EQUAL(2ULL, uni.Count());
+  BOOST_CHECK_EQUAL(2UL, *uni.begin());
+
+  BOOST_REQUIRE_EQUAL(NGram::TotalSize(2) * 4, outputs[1].Size());
+  NGram bi(outputs[1].Get(), 2);
+  BOOST_CHECK_EQUAL(0UL, *bi.begin());
+  BOOST_CHECK_EQUAL(0UL, *(bi.begin() + 1));
+  BOOST_CHECK_EQUAL(1ULL, bi.Count());
+  bi.NextInMemory();
+}
+
+}}} // namespaces
diff --git a/lm/builder/corpus_count.cc b/lm/builder/corpus_count.cc
new file mode 100644
index 000000000..8c3de57dd
--- /dev/null
+++ b/lm/builder/corpus_count.cc
@@ -0,0 +1,223 @@
+#include "lm/builder/corpus_count.hh"
+
+#include "lm/builder/ngram.hh"
+#include "lm/lm_exception.hh"
+#include "lm/word_index.hh"
+#include "util/file.hh"
+#include "util/file_piece.hh"
+#include "util/murmur_hash.hh"
+#include "util/probing_hash_table.hh"
+#include "util/scoped.hh"
+#include "util/stream/chain.hh"
+#include "util/stream/timer.hh"
+#include "util/tokenize_piece.hh"
+
+#include <boost/unordered_set.hpp>
+#include <boost/unordered_map.hpp>
+
+#include <functional>
+
+#include <stdint.h>
+
+namespace lm {
+namespace builder {
+namespace {
+
+class VocabHandout {
+  public:
+    explicit VocabHandout(int fd) {
+      util::scoped_fd duped(util::DupOrThrow(fd));
+      word_list_.reset(util::FDOpenOrThrow(duped));
+      
+      Lookup("<unk>"); // Force 0
+      Lookup("<s>"); // Force 1
+      Lookup("</s>"); // Force 2
+    }
+
+    WordIndex Lookup(const StringPiece &word) {
+      uint64_t hashed = util::MurmurHashNative(word.data(), word.size());
+      std::pair<Seen::iterator, bool> ret(seen_.insert(std::pair<uint64_t, lm::WordIndex>(hashed, seen_.size())));
+      if (ret.second) {
+        char null_delimit = 0;
+        util::WriteOrThrow(word_list_.get(), word.data(), word.size());
+        util::WriteOrThrow(word_list_.get(), &null_delimit, 1);
+        UTIL_THROW_IF(seen_.size() >= std::numeric_limits<lm::WordIndex>::max(), VocabLoadException, "Too many vocabulary words.  Change WordIndex to uint64_t in lm/word_index.hh.");
+      }
+      return ret.first->second;
+    }
+
+    WordIndex Size() const {
+      return seen_.size();
+    }
+
+  private:
+    typedef boost::unordered_map<uint64_t, lm::WordIndex> Seen;
+
+    Seen seen_;
+
+    util::scoped_FILE word_list_;
+};
+
+class DedupeHash : public std::unary_function<const WordIndex *, bool> {
+  public:
+    explicit DedupeHash(std::size_t order) : size_(order * sizeof(WordIndex)) {}
+
+    std::size_t operator()(const WordIndex *start) const {
+      return util::MurmurHashNative(start, size_);
+    }
+    
+  private:
+    const std::size_t size_;
+};
+
+class DedupeEquals : public std::binary_function<const WordIndex *, const WordIndex *, bool> {
+  public:
+    explicit DedupeEquals(std::size_t order) : size_(order * sizeof(WordIndex)) {}
+    
+    bool operator()(const WordIndex *first, const WordIndex *second) const {
+      return !memcmp(first, second, size_);
+    } 
+    
+  private:
+    const std::size_t size_;
+};
+
+struct DedupeEntry {
+  typedef WordIndex *Key;
+  Key GetKey() const { return key; }
+  Key key;
+  static DedupeEntry Construct(WordIndex *at) {
+    DedupeEntry ret;
+    ret.key = at;
+    return ret;
+  }
+};
+
+typedef util::ProbingHashTable<DedupeEntry, DedupeHash, DedupeEquals> Dedupe;
+
+const float kProbingMultiplier = 1.5;
+
+class Writer {
+  public:
+    Writer(std::size_t order, const util::stream::ChainPosition &position, void *dedupe_mem, std::size_t dedupe_mem_size) 
+      : block_(position), gram_(block_->Get(), order),
+        dedupe_invalid_(order, std::numeric_limits<WordIndex>::max()),
+        dedupe_(dedupe_mem, dedupe_mem_size, &dedupe_invalid_[0], DedupeHash(order), DedupeEquals(order)),
+        buffer_(new WordIndex[order - 1]),
+        block_size_(position.GetChain().BlockSize()) {
+      dedupe_.Clear(DedupeEntry::Construct(&dedupe_invalid_[0]));
+      assert(Dedupe::Size(position.GetChain().BlockSize() / position.GetChain().EntrySize(), kProbingMultiplier) == dedupe_mem_size);
+      if (order == 1) {
+        // Add special words.  AdjustCounts is responsible if order != 1.    
+        AddUnigramWord(kUNK);
+        AddUnigramWord(kBOS);
+      }
+    }
+
+    ~Writer() {
+      block_->SetValidSize(reinterpret_cast<const uint8_t*>(gram_.begin()) - static_cast<const uint8_t*>(block_->Get()));
+      (++block_).Poison();
+    }
+
+    // Write context with a bunch of <s>
+    void StartSentence() {
+      for (WordIndex *i = gram_.begin(); i != gram_.end() - 1; ++i) {
+        *i = kBOS;
+      }
+    }
+
+    void Append(WordIndex word) {
+      *(gram_.end() - 1) = word;
+      Dedupe::MutableIterator at;
+      bool found = dedupe_.FindOrInsert(DedupeEntry::Construct(gram_.begin()), at);
+      if (found) {
+        // Already present.
+        NGram already(at->key, gram_.Order());
+        ++(already.Count());
+        // Shift left by one.
+        memmove(gram_.begin(), gram_.begin() + 1, sizeof(WordIndex) * (gram_.Order() - 1));
+        return;
+      }
+      // Complete the write.  
+      gram_.Count() = 1;
+      // Prepare the next n-gram.  
+      if (reinterpret_cast<uint8_t*>(gram_.begin()) + gram_.TotalSize() != static_cast<uint8_t*>(block_->Get()) + block_size_) {
+        NGram last(gram_);
+        gram_.NextInMemory();
+        std::copy(last.begin() + 1, last.end(), gram_.begin());
+        return;
+      }
+      // Block end.  Need to store the context in a temporary buffer.  
+      std::copy(gram_.begin() + 1, gram_.end(), buffer_.get());
+      dedupe_.Clear(DedupeEntry::Construct(&dedupe_invalid_[0]));
+      block_->SetValidSize(block_size_);
+      gram_.ReBase((++block_)->Get());
+      std::copy(buffer_.get(), buffer_.get() + gram_.Order() - 1, gram_.begin());
+    }
+
+  private:
+    void AddUnigramWord(WordIndex index) {
+      *gram_.begin() = index;
+      gram_.Count() = 0;
+      gram_.NextInMemory();
+      if (gram_.Base() == static_cast<uint8_t*>(block_->Get()) + block_size_) {
+        block_->SetValidSize(block_size_);
+        gram_.ReBase((++block_)->Get());
+      }
+    }
+
+    util::stream::Link block_;
+
+    NGram gram_;
+
+    // This is the memory behind the invalid value in dedupe_.
+    std::vector<WordIndex> dedupe_invalid_;
+    // Hash table combiner implementation.
+    Dedupe dedupe_;
+
+    // Small buffer to hold existing ngrams when shifting across a block boundary.  
+    boost::scoped_array<WordIndex> buffer_;
+
+    const std::size_t block_size_;
+};
+
+} // namespace
+
+float CorpusCount::DedupeMultiplier(std::size_t order) {
+  return kProbingMultiplier * static_cast<float>(sizeof(DedupeEntry)) / static_cast<float>(NGram::TotalSize(order));
+}
+
+CorpusCount::CorpusCount(util::FilePiece &from, int vocab_write, uint64_t &token_count, WordIndex &type_count, std::size_t entries_per_block) 
+  : from_(from), vocab_write_(vocab_write), token_count_(token_count), type_count_(type_count),
+    dedupe_mem_size_(Dedupe::Size(entries_per_block, kProbingMultiplier)),
+    dedupe_mem_(util::MallocOrThrow(dedupe_mem_size_)) {
+  token_count_ = 0;
+  type_count_ = 0;
+}
+
+void CorpusCount::Run(const util::stream::ChainPosition &position) {
+  UTIL_TIMER("(%w s) Counted n-grams\n");
+
+  VocabHandout vocab(vocab_write_);
+  const WordIndex end_sentence = vocab.Lookup("</s>");
+  Writer writer(NGram::OrderFromSize(position.GetChain().EntrySize()), position, dedupe_mem_.get(), dedupe_mem_size_);
+  uint64_t count = 0;
+  try {
+    while(true) {
+      StringPiece line(from_.ReadLine());
+      writer.StartSentence();
+      for (util::TokenIter<util::AnyCharacter, true> w(line, " \t"); w; ++w) {
+        WordIndex word = vocab.Lookup(*w);
+        UTIL_THROW_IF(word <= 2, FormatLoadException, "Special word " << *w << " is not allowed in the corpus.  I plan to support models containing <unk> in the future.");
+        writer.Append(word);
+        ++count;
+      }
+      writer.Append(end_sentence);
+    }
+  } catch (const util::EndOfFileException &e) {}
+  token_count_ = count;
+  type_count_ = vocab.Size();
+}
+
+} // namespace builder
+} // namespace lm
diff --git a/lm/builder/corpus_count.hh b/lm/builder/corpus_count.hh
new file mode 100644
index 000000000..e255bad13
--- /dev/null
+++ b/lm/builder/corpus_count.hh
@@ -0,0 +1,42 @@
+#ifndef LM_BUILDER_CORPUS_COUNT__
+#define LM_BUILDER_CORPUS_COUNT__
+
+#include "lm/word_index.hh"
+#include "util/scoped.hh"
+
+#include <cstddef>
+#include <string>
+#include <stdint.h>
+
+namespace util {
+class FilePiece;
+namespace stream {
+class ChainPosition;
+} // namespace stream
+} // namespace util
+
+namespace lm {
+namespace builder {
+
+class CorpusCount {
+  public:
+    // Memory usage will be DedupeMultipler(order) * block_size + total_chain_size + unknown vocab_hash_size
+    static float DedupeMultiplier(std::size_t order);
+
+    CorpusCount(util::FilePiece &from, int vocab_write, uint64_t &token_count, WordIndex &type_count, std::size_t entries_per_block);
+
+    void Run(const util::stream::ChainPosition &position);
+
+  private:
+    util::FilePiece &from_;
+    int vocab_write_;
+    uint64_t &token_count_;
+    WordIndex &type_count_;
+
+    std::size_t dedupe_mem_size_;
+    util::scoped_malloc dedupe_mem_;
+};
+
+} // namespace builder
+} // namespace lm
+#endif // LM_BUILDER_CORPUS_COUNT__
diff --git a/lm/builder/corpus_count_test.cc b/lm/builder/corpus_count_test.cc
new file mode 100644
index 000000000..8d53ca9d1
--- /dev/null
+++ b/lm/builder/corpus_count_test.cc
@@ -0,0 +1,76 @@
+#include "lm/builder/corpus_count.hh"
+
+#include "lm/builder/ngram.hh"
+#include "lm/builder/ngram_stream.hh"
+
+#include "util/file.hh"
+#include "util/file_piece.hh"
+#include "util/tokenize_piece.hh"
+#include "util/stream/chain.hh"
+#include "util/stream/stream.hh"
+
+#define BOOST_TEST_MODULE CorpusCountTest
+#include <boost/test/unit_test.hpp>
+
+namespace lm { namespace builder { namespace {
+
+#define Check(str, count) { \
+  BOOST_REQUIRE(stream); \
+  w = stream->begin(); \
+  for (util::TokenIter<util::AnyCharacter, true> t(str, " "); t; ++t, ++w) { \
+    BOOST_CHECK_EQUAL(*t, v[*w]); \
+  } \
+  BOOST_CHECK_EQUAL((uint64_t)count, stream->Count()); \
+  ++stream; \
+}
+
+BOOST_AUTO_TEST_CASE(Short) {
+  util::scoped_fd input_file(util::MakeTemp("corpus_count_test_temp"));
+  const char input[] = "looking on a little more loin\non a little more loin\non foo little more loin\nbar\n\n";
+  // Blocks of 10 are
+  // looking on a little more loin </s> on a little[duplicate] more[duplicate] loin[duplicate] </s>[duplicate] on[duplicate] foo
+  // little more loin </s> bar </s> </s>
+
+  util::WriteOrThrow(input_file.get(), input, sizeof(input) - 1);
+  util::FilePiece input_piece(input_file.release(), "temp file");
+
+  util::stream::ChainConfig config;
+  config.entry_size = NGram::TotalSize(3);
+  config.total_memory = config.entry_size * 20;
+  config.block_count = 2;
+
+  util::scoped_fd vocab(util::MakeTemp("corpus_count_test_vocab"));
+
+  util::stream::Chain chain(config);
+  NGramStream stream;
+  uint64_t token_count;
+  WordIndex type_count;
+  CorpusCount counter(input_piece, vocab.get(), token_count, type_count, chain.BlockSize() / chain.EntrySize());
+  chain >> boost::ref(counter) >> stream >> util::stream::kRecycle;
+
+  const char *v[] = {"<unk>", "<s>", "</s>", "looking", "on", "a", "little", "more", "loin", "foo", "bar"};
+
+  WordIndex *w;
+
+  Check("<s> <s> looking", 1);
+  Check("<s> looking on", 1);
+  Check("looking on a", 1);
+  Check("on a little", 2);
+  Check("a little more", 2);
+  Check("little more loin", 2);
+  Check("more loin </s>", 2);
+  Check("<s> <s> on", 2);
+  Check("<s> on a", 1);
+  Check("<s> on foo", 1);
+  Check("on foo little", 1);
+  Check("foo little more", 1);
+  Check("little more loin", 1);
+  Check("more loin </s>", 1);
+  Check("<s> <s> bar", 1);
+  Check("<s> bar </s>", 1);
+  Check("<s> <s> </s>", 1);
+  BOOST_CHECK(!stream);
+  BOOST_CHECK_EQUAL(sizeof(v) / sizeof(const char*), type_count);
+}
+
+}}} // namespaces
diff --git a/lm/builder/discount.hh b/lm/builder/discount.hh
new file mode 100644
index 000000000..754fb20db
--- /dev/null
+++ b/lm/builder/discount.hh
@@ -0,0 +1,26 @@
+#ifndef BUILDER_DISCOUNT__
+#define BUILDER_DISCOUNT__
+
+#include <algorithm>
+
+#include <inttypes.h>
+
+namespace lm {
+namespace builder {
+
+struct Discount {
+  float amount[4];
+
+  float Get(uint64_t count) const {
+    return amount[std::min<uint64_t>(count, 3)];
+  }
+
+  float Apply(uint64_t count) const {
+    return static_cast<float>(count) - Get(count);
+  }
+};
+
+} // namespace builder
+} // namespace lm
+
+#endif // BUILDER_DISCOUNT__
diff --git a/lm/builder/header_info.hh b/lm/builder/header_info.hh
new file mode 100644
index 000000000..ccca14566
--- /dev/null
+++ b/lm/builder/header_info.hh
@@ -0,0 +1,20 @@
+#ifndef LM_BUILDER_HEADER_INFO__
+#define LM_BUILDER_HEADER_INFO__
+
+#include <string>
+#include <stdint.h>
+
+// Some configuration info that is used to add
+// comments to the beginning of an ARPA file
+struct HeaderInfo {
+  const std::string input_file;
+  const uint64_t token_count;
+
+  HeaderInfo(const std::string& input_file_in, uint64_t token_count_in)
+    : input_file(input_file_in), token_count(token_count_in) {}
+
+  // TODO: Add smoothing type
+  // TODO: More info if multiple models were interpolated
+};
+
+#endif
diff --git a/lm/builder/initial_probabilities.cc b/lm/builder/initial_probabilities.cc
new file mode 100644
index 000000000..58b42a20c
--- /dev/null
+++ b/lm/builder/initial_probabilities.cc
@@ -0,0 +1,136 @@
+#include "lm/builder/initial_probabilities.hh"
+
+#include "lm/builder/discount.hh"
+#include "lm/builder/ngram_stream.hh"
+#include "lm/builder/sort.hh"
+#include "util/file.hh"
+#include "util/stream/chain.hh"
+#include "util/stream/io.hh"
+#include "util/stream/stream.hh"
+
+#include <vector>
+
+namespace lm { namespace builder {
+
+namespace {
+struct BufferEntry {
+  // Gamma from page 20 of Chen and Goodman.  
+  float gamma;
+  // \sum_w a(c w) for all w.  
+  float denominator;
+};
+
+// Extract an array of gamma from an array of BufferEntry.  
+class OnlyGamma {
+  public:
+    void Run(const util::stream::ChainPosition &position) {
+      for (util::stream::Link block_it(position); block_it; ++block_it) {
+        float *out = static_cast<float*>(block_it->Get());
+        const float *in = out;
+        const float *end = static_cast<const float*>(block_it->ValidEnd());
+        for (out += 1, in += 2; in < end; out += 1, in += 2) {
+          *out = *in;
+        }
+        block_it->SetValidSize(block_it->ValidSize() / 2);
+      }
+    }
+};
+
+class AddRight {
+  public:
+    AddRight(const Discount &discount, const util::stream::ChainPosition &input) 
+      : discount_(discount), input_(input) {}
+
+    void Run(const util::stream::ChainPosition &output) {
+      NGramStream in(input_);
+      util::stream::Stream out(output);
+
+      std::vector<WordIndex> previous(in->Order() - 1);
+      const std::size_t size = sizeof(WordIndex) * previous.size();
+      for(; in; ++out) {
+        memcpy(&previous[0], in->begin(), size);
+        uint64_t denominator = 0;
+        uint64_t counts[4];
+        memset(counts, 0, sizeof(counts));
+        do {
+          denominator += in->Count();
+          ++counts[std::min(in->Count(), static_cast<uint64_t>(3))];
+        } while (++in && !memcmp(&previous[0], in->begin(), size));
+        BufferEntry &entry = *reinterpret_cast<BufferEntry*>(out.Get());
+        entry.denominator = static_cast<float>(denominator);
+        entry.gamma = 0.0;
+        for (unsigned i = 1; i <= 3; ++i) {
+          entry.gamma += discount_.Get(i) * static_cast<float>(counts[i]);
+        }
+        entry.gamma /= entry.denominator;
+      }
+      out.Poison();
+    }
+
+  private:
+    const Discount &discount_;
+    const util::stream::ChainPosition input_;
+};
+
+class MergeRight {
+  public:
+    MergeRight(bool interpolate_unigrams, const util::stream::ChainPosition &from_adder, const Discount &discount)
+      : interpolate_unigrams_(interpolate_unigrams), from_adder_(from_adder), discount_(discount) {}
+
+    // calculate the initial probability of each n-gram (before order-interpolation)
+    // Run() gets invoked once for each order
+    void Run(const util::stream::ChainPosition &primary) {
+      util::stream::Stream summed(from_adder_);
+
+      NGramStream grams(primary);
+
+      // Without interpolation, the interpolation weight goes to <unk>.
+      if (grams->Order() == 1 && !interpolate_unigrams_) {
+        BufferEntry sums(*static_cast<const BufferEntry*>(summed.Get()));
+        assert(*grams->begin() == kUNK);
+        grams->Value().uninterp.prob = sums.gamma;
+        grams->Value().uninterp.gamma = 0.0;
+        while (++grams) {
+          grams->Value().uninterp.prob = discount_.Apply(grams->Count()) / sums.denominator;
+          grams->Value().uninterp.gamma = 0.0;
+        }
+        ++summed;
+        return;
+      }
+
+      std::vector<WordIndex> previous(grams->Order() - 1);
+      const std::size_t size = sizeof(WordIndex) * previous.size();
+      for (; grams; ++summed) {
+        memcpy(&previous[0], grams->begin(), size);
+        const BufferEntry &sums = *static_cast<const BufferEntry*>(summed.Get());
+        do {
+          Payload &pay = grams->Value();
+          pay.uninterp.prob = discount_.Apply(pay.count) / sums.denominator;
+          pay.uninterp.gamma = sums.gamma;
+        } while (++grams && !memcmp(&previous[0], grams->begin(), size));
+      }
+    }
+
+  private:
+    bool interpolate_unigrams_;
+    util::stream::ChainPosition from_adder_;
+    Discount discount_;
+};
+
+} // namespace
+
+void InitialProbabilities(const InitialProbabilitiesConfig &config, const std::vector<Discount> &discounts, Chains &primary, Chains &second_in, Chains &gamma_out) {
+  util::stream::ChainConfig gamma_config = config.adder_out;
+  gamma_config.entry_size = sizeof(BufferEntry);
+  for (size_t i = 0; i < primary.size(); ++i) {
+    util::stream::ChainPosition second(second_in[i].Add());
+    second_in[i] >> util::stream::kRecycle;
+    gamma_out.push_back(gamma_config);
+    gamma_out[i] >> AddRight(discounts[i], second);
+    primary[i] >> MergeRight(config.interpolate_unigrams, gamma_out[i].Add(), discounts[i]);
+    // Don't bother with the OnlyGamma thread for something to discard.  
+    if (i) gamma_out[i] >> OnlyGamma();
+  }
+}
+
+}} // namespaces
diff --git a/lm/builder/initial_probabilities.hh b/lm/builder/initial_probabilities.hh
new file mode 100644
index 000000000..626388eb0
--- /dev/null
+++ b/lm/builder/initial_probabilities.hh
@@ -0,0 +1,34 @@
+#ifndef LM_BUILDER_INITIAL_PROBABILITIES__
+#define LM_BUILDER_INITIAL_PROBABILITIES__
+
+#include "lm/builder/discount.hh"
+#include "util/stream/config.hh"
+
+#include <vector>
+
+namespace lm {
+namespace builder {
+class Chains;
+
+struct InitialProbabilitiesConfig {
+  // These should be small buffers to keep the adder from getting too far ahead
+  util::stream::ChainConfig adder_in;
+  util::stream::ChainConfig adder_out;
+  // SRILM doesn't normally interpolate unigrams.  
+  bool interpolate_unigrams;
+};
+
+/* Compute initial (uninterpolated) probabilities
+ * primary: the normal chain of n-grams.  Incoming is context sorted adjusted
+ *   counts.  Outgoing has uninterpolated probabilities for use by Interpolate.
+ * second_in: a second copy of the primary input.  Discard the output.  
+ * gamma_out: Computed gamma values are output on these chains in suffix order.
+ *   The values are bare floats and should be buffered for interpolation to
+ *   use.  
+ */
+void InitialProbabilities(const InitialProbabilitiesConfig &config, const std::vector<Discount> &discounts, Chains &primary, Chains &second_in, Chains &gamma_out);
+
+} // namespace builder
+} // namespace lm
+
+#endif // LM_BUILDER_INITIAL_PROBABILITIES__
diff --git a/lm/builder/interpolate.cc b/lm/builder/interpolate.cc
new file mode 100644
index 000000000..500268069
--- /dev/null
+++ b/lm/builder/interpolate.cc
@@ -0,0 +1,65 @@
+#include "lm/builder/interpolate.hh"
+
+#include "lm/builder/joint_order.hh"
+#include "lm/builder/multi_stream.hh"
+#include "lm/builder/sort.hh"
+#include "lm/lm_exception.hh"
+
+#include <assert.h>
+
+namespace lm { namespace builder {
+namespace {
+
+class Callback {
+  public:
+    Callback(float uniform_prob, const ChainPositions &backoffs) : backoffs_(backoffs.size()), probs_(backoffs.size() + 2) {
+      probs_[0] = uniform_prob;
+      for (std::size_t i = 0; i < backoffs.size(); ++i) {
+        backoffs_.push_back(backoffs[i]);
+      }
+    }
+
+    ~Callback() {
+      for (std::size_t i = 0; i < backoffs_.size(); ++i) {
+        if (backoffs_[i]) {
+          std::cerr << "Backoffs do not match for order " << (i + 1) << std::endl;
+          abort();
+        }
+      }
+    }
+
+    void Enter(unsigned order_minus_1, NGram &gram) {
+      Payload &pay = gram.Value();
+      pay.complete.prob = pay.uninterp.prob + pay.uninterp.gamma * probs_[order_minus_1];
+      probs_[order_minus_1 + 1] = pay.complete.prob;
+      pay.complete.prob = log10(pay.complete.prob);
+      // TODO: this is a hack to skip n-grams that don't appear as context.  Pruning will require some different handling.  
+      if (order_minus_1 < backoffs_.size() && *(gram.end() - 1) != kUNK && *(gram.end() - 1) != kEOS) {
+        pay.complete.backoff = log10(*static_cast<const float*>(backoffs_[order_minus_1].Get()));
+        ++backoffs_[order_minus_1];
+      } else {
+        // Not a context.  
+        pay.complete.backoff = 0.0;
+      }
+    }
+
+    void Exit(unsigned, const NGram &) const {}
+
+  private:
+    FixedArray<util::stream::Stream> backoffs_;
+
+    std::vector<float> probs_;
+};
+} // namespace
+
+Interpolate::Interpolate(uint64_t unigram_count, const ChainPositions &backoffs) 
+  : uniform_prob_(1.0 / static_cast<float>(unigram_count - 1)), backoffs_(backoffs) {}
+
+// perform order-wise interpolation
+void Interpolate::Run(const ChainPositions &positions) {
+  assert(positions.size() == backoffs_.size() + 1);
+  Callback callback(uniform_prob_, backoffs_);
+  JointOrder<Callback, SuffixOrder>(positions, callback);
+}
+
+}} // namespaces
diff --git a/lm/builder/interpolate.hh b/lm/builder/interpolate.hh
new file mode 100644
index 000000000..9268d404c
--- /dev/null
+++ b/lm/builder/interpolate.hh
@@ -0,0 +1,27 @@
+#ifndef LM_BUILDER_INTERPOLATE__
+#define LM_BUILDER_INTERPOLATE__
+
+#include <stdint.h>
+
+#include "lm/builder/multi_stream.hh"
+
+namespace lm { namespace builder {
+ 
+/* Interpolate step.  
+ * Input: suffix sorted n-grams with (p_uninterpolated, gamma) from
+ * InitialProbabilities.
+ * Output: suffix sorted n-grams with complete probability
+ */
+class Interpolate {
+  public:
+    explicit Interpolate(uint64_t unigram_count, const ChainPositions &backoffs);
+
+    void Run(const ChainPositions &positions);
+
+  private:
+    float uniform_prob_;
+    ChainPositions backoffs_;
+};
+
+}} // namespaces
+#endif // LM_BUILDER_INTERPOLATE__
diff --git a/lm/builder/joint_order.hh b/lm/builder/joint_order.hh
new file mode 100644
index 000000000..b56201447
--- /dev/null
+++ b/lm/builder/joint_order.hh
@@ -0,0 +1,43 @@
+#ifndef LM_BUILDER_JOINT_ORDER__
+#define LM_BUILDER_JOINT_ORDER__
+
+#include "lm/builder/multi_stream.hh"
+#include "lm/lm_exception.hh"
+
+#include <string.h>
+
+namespace lm { namespace builder {
+
+template <class Callback, class Compare> void JointOrder(const ChainPositions &positions, Callback &callback) {
+  // Allow matching to reference streams[-1].
+  NGramStreams streams_with_dummy;
+  streams_with_dummy.InitWithDummy(positions);
+  NGramStream *streams = streams_with_dummy.begin() + 1;
+
+  unsigned int order;
+  for (order = 0; order < positions.size() && streams[order]; ++order) {}
+  assert(order); // should always have <unk>.
+  unsigned int current = 0;
+  while (true) {
+    // Does the context match the lower one?  
+    if (!memcmp(streams[static_cast<int>(current) - 1]->begin(), streams[current]->begin() + Compare::kMatchOffset, sizeof(WordIndex) * current)) {
+      callback.Enter(current, *streams[current]);
+      // Transition to looking for extensions.  
+      if (++current < order) continue;
+    }
+    // No extension left.  
+    while(true) {
+      assert(current > 0);
+      --current;
+      callback.Exit(current, *streams[current]);
+      if (++streams[current]) break;
+      UTIL_THROW_IF(order != current + 1, FormatLoadException, "Detected n-gram without matching suffix");
+      order = current;
+      if (!order) return;
+    }
+  }
+}
+
+}} // namespaces
+
+#endif // LM_BUILDER_JOINT_ORDER__
diff --git a/lm/builder/main.cc b/lm/builder/main.cc
new file mode 100644
index 000000000..90b9dca2d
--- /dev/null
+++ b/lm/builder/main.cc
@@ -0,0 +1,94 @@
+#include "lm/builder/pipeline.hh"
+#include "util/file.hh"
+#include "util/file_piece.hh"
+#include "util/usage.hh"
+
+#include <iostream>
+
+#include <boost/program_options.hpp>
+
+namespace {
+class SizeNotify {
+  public:
+    SizeNotify(std::size_t &out) : behind_(out) {}
+
+    void operator()(const std::string &from) {
+      behind_ = util::ParseSize(from);
+    }
+
+  private:
+    std::size_t &behind_;
+};
+
+boost::program_options::typed_value<std::string> *SizeOption(std::size_t &to, const char *default_value) {
+  return boost::program_options::value<std::string>()->notifier(SizeNotify(to))->default_value(default_value);
+}
+
+} // namespace
+
+int main(int argc, char *argv[]) {
+  try {
+    namespace po = boost::program_options;
+    po::options_description options("Language model building options");
+    lm::builder::PipelineConfig pipeline;
+
+    options.add_options()
+      ("order,o", po::value<std::size_t>(&pipeline.order)->required(), "Order of the model")
+      ("interpolate_unigrams", po::bool_switch(&pipeline.initial_probs.interpolate_unigrams), "Interpolate the unigrams (default: emulate SRILM by not interpolating)")
+      ("temp_prefix,T", po::value<std::string>(&pipeline.sort.temp_prefix)->default_value("/tmp/lm"), "Temporary file prefix")
+      ("memory,S", SizeOption(pipeline.sort.total_memory, util::GuessPhysicalMemory() ? "80%" : "1G"), "Sorting memory")
+      ("vocab_memory", SizeOption(pipeline.assume_vocab_hash_size, "50M"), "Assume that the vocabulary hash table will use this much memory for purposes of calculating total memory in the count step")
+      ("minimum_block", SizeOption(pipeline.minimum_block, "8K"), "Minimum block size to allow")
+      ("sort_block", SizeOption(pipeline.sort.buffer_size, "64M"), "Size of IO operations for sort (determines arity)")
+      ("block_count", po::value<std::size_t>(&pipeline.block_count)->default_value(2), "Block count (per order)")
+      ("vocab_file", po::value<std::string>(&pipeline.vocab_file)->default_value(""), "Location to write vocabulary file")
+      ("verbose_header", po::bool_switch(&pipeline.verbose_header), "Add a verbose header to the ARPA file that includes information such as token count, smoothing type, etc.");
+    if (argc == 1) {
+      std::cerr << 
+        "Builds unpruned language models with modified Kneser-Ney smoothing.\n\n"
+        "Please cite:\n"
+        "@inproceedings{kenlm,\n"
+        "author    = {Kenneth Heafield},\n"
+        "title     = {{KenLM}: Faster and Smaller Language Model Queries},\n"
+        "booktitle = {Proceedings of the Sixth Workshop on Statistical Machine Translation},\n"
+        "month     = {July}, year={2011},\n"
+        "address   = {Edinburgh, UK},\n"
+        "publisher = {Association for Computational Linguistics},\n"
+        "}\n\n"
+        "Provide the corpus on stdin.  The ARPA file will be written to stdout.  Order of\n"
+        "the model (-o) is the only mandatory option.  As this is an on-disk program,\n"
+        "setting the temporary file location (-T) and sorting memory (-S) is recommended.\n\n"
+        "Memory sizes are specified like GNU sort: a number followed by a unit character.\n"
+        "Valid units are \% for percentage of memory (supported platforms only) and (in\n"
+        "increasing powers of 1024): b, K, M, G, T, P, E, Z, Y.  Default is K (*1024).\n\n";
+      std::cerr << options << std::endl;
+      return 1;
+    }
+    po::variables_map vm;
+    po::store(po::parse_command_line(argc, argv, options), vm);
+    po::notify(vm);
+
+    util::NormalizeTempPrefix(pipeline.sort.temp_prefix);
+
+    lm::builder::InitialProbabilitiesConfig &initial = pipeline.initial_probs;
+    // TODO: evaluate options for these.  
+    initial.adder_in.total_memory = 32768;
+    initial.adder_in.block_count = 2;
+    initial.adder_out.total_memory = 32768;
+    initial.adder_out.block_count = 2;
+    pipeline.read_backoffs = initial.adder_out;
+
+    // Read from stdin
+    try {
+      lm::builder::Pipeline(pipeline, 0, 1);
+    } catch (const util::MallocException &e) {
+      std::cerr << e.what() << std::endl;
+      std::cerr << "Try rerunning with a more conservative -S setting than " << vm["memory"].as<std::string>() << std::endl;
+      return 1;
+    }
+    util::PrintUsage(std::cerr);
+  } catch (const std::exception &e) {
+    std::cerr << e.what() << std::endl;
+    return 1;
+  }
+}
diff --git a/lm/builder/multi_stream.hh b/lm/builder/multi_stream.hh
new file mode 100644
index 000000000..707a98c7f
--- /dev/null
+++ b/lm/builder/multi_stream.hh
@@ -0,0 +1,180 @@
+#ifndef LM_BUILDER_MULTI_STREAM__
+#define LM_BUILDER_MULTI_STREAM__
+
+#include "lm/builder/ngram_stream.hh"
+#include "util/scoped.hh"
+#include "util/stream/chain.hh"
+
+#include <cstddef>
+#include <new>
+
+#include <assert.h>
+#include <stdlib.h>
+
+namespace lm { namespace builder {
+
+template <class T> class FixedArray {
+  public:
+    explicit FixedArray(std::size_t count) {
+      Init(count);
+    }
+
+    FixedArray() : newed_end_(NULL) {}
+
+    void Init(std::size_t count) {
+      assert(!block_.get());
+      block_.reset(malloc(sizeof(T) * count));
+      if (!block_.get()) throw std::bad_alloc();
+      newed_end_ = begin();
+    }
+
+    FixedArray(const FixedArray &from) {
+      std::size_t size = from.newed_end_ - static_cast<const T*>(from.block_.get());
+      Init(size);
+      for (std::size_t i = 0; i < size; ++i) {
+        new(end()) T(from[i]);
+        Constructed();
+      }
+    }
+
+    ~FixedArray() { clear(); }
+
+    T *begin() { return static_cast<T*>(block_.get()); }
+    const T *begin() const { return static_cast<const T*>(block_.get()); }
+    // Always call Constructed after successful completion of new.  
+    T *end() { return newed_end_; }
+    const T *end() const { return newed_end_; }
+
+    T &back() { return *(end() - 1); }
+    const T &back() const { return *(end() - 1); }
+
+    std::size_t size() const { return end() - begin(); }
+    bool empty() const { return begin() == end(); }
+
+    T &operator[](std::size_t i) { return begin()[i]; }
+    const T &operator[](std::size_t i) const { return begin()[i]; }
+
+    template <class C> void push_back(const C &c) {
+      new (end()) T(c);
+      Constructed();
+    }
+
+    void clear() {
+      for (T *i = begin(); i != end(); ++i)
+        i->~T();
+      newed_end_ = begin();
+    }
+
+  protected:
+    void Constructed() {
+      ++newed_end_;
+    }
+
+  private:
+    util::scoped_malloc block_;
+
+    T *newed_end_;
+};
+
+class Chains;
+
+class ChainPositions : public FixedArray<util::stream::ChainPosition> {
+  public:
+    ChainPositions() {}
+
+    void Init(Chains &chains);
+
+    explicit ChainPositions(Chains &chains) {
+      Init(chains);
+    }
+};
+
+class Chains : public FixedArray<util::stream::Chain> {
+  private:
+    template <class T, void (T::*ptr)(const ChainPositions &) = &T::Run> struct CheckForRun {
+      typedef Chains type;
+    };
+
+  public:
+    explicit Chains(std::size_t limit) : FixedArray<util::stream::Chain>(limit) {}
+
+    template <class Worker> typename CheckForRun<Worker>::type &operator>>(const Worker &worker) {
+      threads_.push_back(new util::stream::Thread(ChainPositions(*this), worker));
+      return *this;
+    }
+
+    template <class Worker> typename CheckForRun<Worker>::type &operator>>(const boost::reference_wrapper<Worker> &worker) {
+      threads_.push_back(new util::stream::Thread(ChainPositions(*this), worker));
+      return *this;
+    }
+
+    Chains &operator>>(const util::stream::Recycler &recycler) {
+      for (util::stream::Chain *i = begin(); i != end(); ++i) 
+        *i >> recycler;
+      return *this;
+    }
+
+    void Wait(bool release_memory = true) {
+      threads_.clear();
+      for (util::stream::Chain *i = begin(); i != end(); ++i) {
+        i->Wait(release_memory);
+      }
+    }
+
+  private:
+    boost::ptr_vector<util::stream::Thread> threads_;
+
+    Chains(const Chains &);
+    void operator=(const Chains &);
+};
+
+inline void ChainPositions::Init(Chains &chains) {
+  FixedArray<util::stream::ChainPosition>::Init(chains.size());
+  for (util::stream::Chain *i = chains.begin(); i != chains.end(); ++i) {
+    new (end()) util::stream::ChainPosition(i->Add()); Constructed();
+  }
+}
+
+inline Chains &operator>>(Chains &chains, ChainPositions &positions) {
+  positions.Init(chains);
+  return chains;
+}
+
+class NGramStreams : public FixedArray<NGramStream> {
+  public:
+    NGramStreams() {}
+
+    // This puts a dummy NGramStream at the beginning (useful to algorithms that need to reference something at the beginning).
+    void InitWithDummy(const ChainPositions &positions) {
+      FixedArray<NGramStream>::Init(positions.size() + 1);
+      new (end()) NGramStream(); Constructed();
+      for (const util::stream::ChainPosition *i = positions.begin(); i != positions.end(); ++i) {
+        push_back(*i);
+      }
+    }
+
+    // Limit restricts to positions[0,limit)
+    void Init(const ChainPositions &positions, std::size_t limit) {
+      FixedArray<NGramStream>::Init(limit);
+      for (const util::stream::ChainPosition *i = positions.begin(); i != positions.begin() + limit; ++i) {
+        push_back(*i);
+      }
+    }
+    void Init(const ChainPositions &positions) {
+      Init(positions, positions.size());
+    }
+
+    NGramStreams(const ChainPositions &positions) {
+      Init(positions);
+    }
+};
+
+inline Chains &operator>>(Chains &chains, NGramStreams &streams) {
+  ChainPositions positions;
+  chains >> positions;
+  streams.Init(positions);
+  return chains;
+}
+
+}} // namespaces
+#endif // LM_BUILDER_MULTI_STREAM__
diff --git a/lm/builder/ngram.hh b/lm/builder/ngram.hh
new file mode 100644
index 000000000..2984ed0b6
--- /dev/null
+++ b/lm/builder/ngram.hh
@@ -0,0 +1,84 @@
+#ifndef LM_BUILDER_NGRAM__
+#define LM_BUILDER_NGRAM__
+
+#include "lm/weights.hh"
+#include "lm/word_index.hh"
+
+#include <cstddef>
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+
+namespace lm {
+namespace builder {
+
+struct Uninterpolated {
+  float prob;  // Uninterpolated probability.
+  float gamma; // Interpolation weight for lower order.
+};
+
+union Payload {
+  uint64_t count;
+  Uninterpolated uninterp;
+  ProbBackoff complete;
+};
+
+class NGram {
+  public:
+    NGram(void *begin, std::size_t order) 
+      : begin_(static_cast<WordIndex*>(begin)), end_(begin_ + order) {}
+
+    const uint8_t *Base() const { return reinterpret_cast<const uint8_t*>(begin_); }
+    uint8_t *Base() { return reinterpret_cast<uint8_t*>(begin_); }
+
+    void ReBase(void *to) {
+      std::size_t difference = end_ - begin_;
+      begin_ = reinterpret_cast<WordIndex*>(to);
+      end_ = begin_ + difference;
+    }
+
+    // Would do operator++ but that can get confusing for a stream.  
+    void NextInMemory() {
+      ReBase(&Value() + 1);
+    }
+
+    // Lower-case in deference to STL.  
+    const WordIndex *begin() const { return begin_; }
+    WordIndex *begin() { return begin_; }
+    const WordIndex *end() const { return end_; }
+    WordIndex *end() { return end_; }
+
+    const Payload &Value() const { return *reinterpret_cast<const Payload *>(end_); }
+    Payload &Value() { return *reinterpret_cast<Payload *>(end_); }
+
+    uint64_t &Count() { return Value().count; }
+    const uint64_t Count() const { return Value().count; }
+
+    std::size_t Order() const { return end_ - begin_; }
+
+    static std::size_t TotalSize(std::size_t order) {
+      return order * sizeof(WordIndex) + sizeof(Payload);
+    }
+    std::size_t TotalSize() const {
+      // Compiler should optimize this.  
+      return TotalSize(Order());
+    }
+    static std::size_t OrderFromSize(std::size_t size) {
+      std::size_t ret = (size - sizeof(Payload)) / sizeof(WordIndex);
+      assert(size == TotalSize(ret));
+      return ret;
+    }
+
+  private:
+    WordIndex *begin_, *end_;
+};
+
+const WordIndex kUNK = 0;
+const WordIndex kBOS = 1;
+const WordIndex kEOS = 2;
+
+} // namespace builder
+} // namespace lm
+
+#endif // LM_BUILDER_NGRAM__
diff --git a/lm/builder/ngram_stream.hh b/lm/builder/ngram_stream.hh
new file mode 100644
index 000000000..3c9946643
--- /dev/null
+++ b/lm/builder/ngram_stream.hh
@@ -0,0 +1,55 @@
+#ifndef LM_BUILDER_NGRAM_STREAM__
+#define LM_BUILDER_NGRAM_STREAM__
+
+#include "lm/builder/ngram.hh"
+#include "util/stream/chain.hh"
+#include "util/stream/stream.hh"
+
+#include <cstddef>
+
+namespace lm { namespace builder {
+
+class NGramStream {
+  public:
+    NGramStream() : gram_(NULL, 0) {}
+
+    NGramStream(const util::stream::ChainPosition &position) : gram_(NULL, 0) {
+      Init(position);
+    }
+
+    void Init(const util::stream::ChainPosition &position) {
+      stream_.Init(position);
+      gram_ = NGram(stream_.Get(), NGram::OrderFromSize(position.GetChain().EntrySize()));
+    }
+
+    NGram &operator*() { return gram_; }
+    const NGram &operator*() const { return gram_; }
+
+    NGram *operator->() { return &gram_; }
+    const NGram *operator->() const { return &gram_; }
+
+    void *Get() { return stream_.Get(); }
+    const void *Get() const { return stream_.Get(); }
+
+    operator bool() const { return stream_; }
+    bool operator!() const { return !stream_; }
+    void Poison() { stream_.Poison(); }
+
+    NGramStream &operator++() {
+      ++stream_;
+      gram_.ReBase(stream_.Get());
+      return *this;
+    }
+
+  private:
+    NGram gram_;
+    util::stream::Stream stream_;
+};
+
+inline util::stream::Chain &operator>>(util::stream::Chain &chain, NGramStream &str) {
+  str.Init(chain.Add());
+  return chain;
+}
+
+}} // namespaces
+#endif // LM_BUILDER_NGRAM_STREAM__
diff --git a/lm/builder/pipeline.cc b/lm/builder/pipeline.cc
new file mode 100644
index 000000000..14a1f7218
--- /dev/null
+++ b/lm/builder/pipeline.cc
@@ -0,0 +1,320 @@
+#include "lm/builder/pipeline.hh"
+
+#include "lm/builder/adjust_counts.hh"
+#include "lm/builder/corpus_count.hh"
+#include "lm/builder/initial_probabilities.hh"
+#include "lm/builder/interpolate.hh"
+#include "lm/builder/print.hh"
+#include "lm/builder/sort.hh"
+
+#include "lm/sizes.hh"
+
+#include "util/exception.hh"
+#include "util/file.hh"
+#include "util/stream/io.hh"
+
+#include <algorithm>
+#include <iostream>
+#include <vector>
+
+namespace lm { namespace builder {
+
+namespace {
+void PrintStatistics(const std::vector<uint64_t> &counts, const std::vector<Discount> &discounts) {
+  std::cerr << "Statistics:\n";
+  for (size_t i = 0; i < counts.size(); ++i) {
+    std::cerr << (i + 1) << ' ' << counts[i];
+    for (size_t d = 1; d <= 3; ++d)
+      std::cerr << " D" << d << (d == 3 ? "+=" : "=") << discounts[i].amount[d];
+    std::cerr << '\n';
+  }
+}
+
+class Master {
+  public:
+    explicit Master(const PipelineConfig &config) 
+      : config_(config), chains_(config.order), files_(config.order) {
+      config_.minimum_block = std::max(NGram::TotalSize(config_.order), config_.minimum_block);
+    }
+
+    const PipelineConfig &Config() const { return config_; }
+
+    Chains &MutableChains() { return chains_; }
+
+    template <class T> Master &operator>>(const T &worker) {
+      chains_ >> worker;
+      return *this;
+    }
+
+    // This takes the (partially) sorted ngrams and sets up for adjusted counts.
+    void InitForAdjust(util::stream::Sort<SuffixOrder, AddCombiner> &ngrams, WordIndex types) {
+      const std::size_t each_order_min = config_.minimum_block * config_.block_count;
+      // We know how many unigrams there are.  Don't allocate more than needed to them.
+      const std::size_t min_chains = (config_.order - 1) * each_order_min +
+        std::min(types * NGram::TotalSize(1), each_order_min);
+      // Do merge sort with calculated laziness.
+      const std::size_t merge_using = ngrams.Merge(std::min(config_.TotalMemory() - min_chains, ngrams.DefaultLazy()));
+
+      std::vector<uint64_t> count_bounds(1, types);
+      CreateChains(config_.TotalMemory() - merge_using, count_bounds);
+      ngrams.Output(chains_.back(), merge_using);
+
+      // Setup unigram file.  
+      files_.push_back(util::MakeTemp(config_.TempPrefix()));
+    }
+
+    // For initial probabilities, but this is generic.
+    void SortAndReadTwice(const std::vector<uint64_t> &counts, Sorts<ContextOrder> &sorts, Chains &second, util::stream::ChainConfig second_config) {
+      // Do merge first before allocating chain memory.
+      for (std::size_t i = 1; i < config_.order; ++i) {
+        sorts[i - 1].Merge(0);
+      }
+      // There's no lazy merge, so just divide memory amongst the chains.
+      CreateChains(config_.TotalMemory(), counts);
+      chains_.back().ActivateProgress();
+      chains_[0] >> files_[0].Source();
+      second_config.entry_size = NGram::TotalSize(1);
+      second.push_back(second_config);
+      second.back() >> files_[0].Source();
+      for (std::size_t i = 1; i < config_.order; ++i) {
+        util::scoped_fd fd(sorts[i - 1].StealCompleted());
+        chains_[i].SetProgressTarget(util::SizeOrThrow(fd.get()));
+        chains_[i] >> util::stream::PRead(util::DupOrThrow(fd.get()), true);
+        second_config.entry_size = NGram::TotalSize(i + 1);
+        second.push_back(second_config);
+        second.back() >> util::stream::PRead(fd.release(), true);
+      }
+    }
+
+    // There is no sort after this, so go for broke on lazy merging.
+    template <class Compare> void MaximumLazyInput(const std::vector<uint64_t> &counts, Sorts<Compare> &sorts) {
+      // Determine the minimum we can use for all the chains.
+      std::size_t min_chains = 0;
+      for (std::size_t i = 0; i < config_.order; ++i) {
+        min_chains += std::min(counts[i] * NGram::TotalSize(i + 1), static_cast<uint64_t>(config_.minimum_block));
+      }
+      std::size_t for_merge = min_chains > config_.TotalMemory() ? 0 : (config_.TotalMemory() - min_chains);
+      std::vector<std::size_t> laziness;
+      // Prioritize longer n-grams.
+      for (util::stream::Sort<SuffixOrder> *i = sorts.end() - 1; i >= sorts.begin(); --i) {
+        laziness.push_back(i->Merge(for_merge));
+        assert(for_merge >= laziness.back());
+        for_merge -= laziness.back();
+      }
+      std::reverse(laziness.begin(), laziness.end());
+
+      CreateChains(for_merge + min_chains, counts);
+      chains_.back().ActivateProgress();
+      chains_[0] >> files_[0].Source();
+      for (std::size_t i = 1; i < config_.order; ++i) {
+        sorts[i - 1].Output(chains_[i], laziness[i - 1]);
+      }
+    }
+
+    void BufferFinal(const std::vector<uint64_t> &counts) {
+      chains_[0] >> files_[0].Sink();
+      for (std::size_t i = 1; i < config_.order; ++i) {
+        files_.push_back(util::MakeTemp(config_.TempPrefix()));
+        chains_[i] >> files_[i].Sink();
+      }
+      chains_.Wait(true);
+      // Use less memory.  Because we can.
+      CreateChains(std::min(config_.sort.buffer_size * config_.order, config_.TotalMemory()), counts);
+      for (std::size_t i = 0; i < config_.order; ++i) {
+        chains_[i] >> files_[i].Source();
+      }
+    }
+
+    template <class Compare> void SetupSorts(Sorts<Compare> &sorts) {
+      sorts.Init(config_.order - 1);
+      // Unigrams don't get sorted because their order is always the same.
+      chains_[0] >> files_[0].Sink();
+      for (std::size_t i = 1; i < config_.order; ++i) {
+        sorts.push_back(chains_[i], config_.sort, Compare(i + 1));
+      }
+      chains_.Wait(true);
+    }
+
+  private:
+    // Create chains, allocating memory to them.  Totally heuristic.  Count
+    // bounds are upper bounds on the counts or not present.
+    void CreateChains(std::size_t remaining_mem, const std::vector<uint64_t> &count_bounds) {
+      std::vector<std::size_t> assignments;
+      assignments.reserve(config_.order);
+      // Start by assigning maximum memory usage (to be refined later).
+      for (std::size_t i = 0; i < count_bounds.size(); ++i) {
+        assignments.push_back(static_cast<std::size_t>(std::min(
+            static_cast<uint64_t>(remaining_mem),
+            count_bounds[i] * static_cast<uint64_t>(NGram::TotalSize(i + 1)))));
+      }
+      assignments.resize(config_.order, remaining_mem);
+
+      // Now we know how much memory everybody wants.  How much will they get?
+      // Proportional to this.
+      std::vector<float> portions;
+      // Indices of orders that have yet to be assigned.
+      std::vector<std::size_t> unassigned;
+      for (std::size_t i = 0; i < config_.order; ++i) {
+        portions.push_back(static_cast<float>((i+1) * NGram::TotalSize(i+1)));
+        unassigned.push_back(i);
+      }
+      /*If somebody doesn't eat their full dinner, give it to the rest of the
+       * family.  Then somebody else might not eat their full dinner etc.  Ends
+       * when everybody unassigned is hungry.
+       */
+      float sum;
+      bool found_more;
+      std::vector<std::size_t> block_count(config_.order);
+      do {
+        sum = 0.0;
+        for (std::size_t i = 0; i < unassigned.size(); ++i) {
+          sum += portions[unassigned[i]];
+        }
+        found_more = false;
+        // If the proportional assignment is more than needed, give it just what it needs.
+        for (std::vector<std::size_t>::iterator i = unassigned.begin(); i != unassigned.end();) {
+          if (assignments[*i] <= remaining_mem * (portions[*i] / sum)) {
+            remaining_mem -= assignments[*i];
+            block_count[*i] = 1;
+            i = unassigned.erase(i);
+            found_more = true;
+          } else {
+            ++i;
+          }
+        }
+      } while (found_more);
+      for (std::vector<std::size_t>::iterator i = unassigned.begin(); i != unassigned.end(); ++i) {
+        assignments[*i] = remaining_mem * (portions[*i] / sum);
+        block_count[*i] = config_.block_count;
+      }
+      chains_.clear();
+      std::cerr << "Chain sizes:";
+      for (std::size_t i = 0; i < config_.order; ++i) {
+        std::cerr << ' ' << (i+1) << ":" << assignments[i];
+        chains_.push_back(util::stream::ChainConfig(NGram::TotalSize(i + 1), block_count[i], assignments[i]));
+      }
+      std::cerr << std::endl;
+    }
+
+    PipelineConfig config_;
+
+    Chains chains_;
+    // Often only unigrams, but sometimes all orders.  
+    FixedArray<util::stream::FileBuffer> files_;
+};
+
+void CountText(int text_file /* input */, int vocab_file /* output */, Master &master, uint64_t &token_count, std::string &text_file_name) {
+  const PipelineConfig &config = master.Config();
+  std::cerr << "=== 1/5 Counting and sorting n-grams ===" << std::endl;
+
+  UTIL_THROW_IF(config.TotalMemory() < config.assume_vocab_hash_size, util::Exception, "Vocab hash size estimate " << config.assume_vocab_hash_size << " exceeds total memory " << config.TotalMemory());
+  std::size_t memory_for_chain = 
+    // This much memory to work with after vocab hash table.
+    static_cast<float>(config.TotalMemory() - config.assume_vocab_hash_size) /
+    // Solve for block size including the dedupe multiplier for one block.
+    (static_cast<float>(config.block_count) + CorpusCount::DedupeMultiplier(config.order)) *
+    // Chain likes memory expressed in terms of total memory.
+    static_cast<float>(config.block_count);
+  util::stream::Chain chain(util::stream::ChainConfig(NGram::TotalSize(config.order), config.block_count, memory_for_chain));
+
+  WordIndex type_count;
+  util::FilePiece text(text_file, NULL, &std::cerr);
+  text_file_name = text.FileName();
+  CorpusCount counter(text, vocab_file, token_count, type_count, chain.BlockSize() / chain.EntrySize());
+  chain >> boost::ref(counter);
+
+  util::stream::Sort<SuffixOrder, AddCombiner> sorter(chain, config.sort, SuffixOrder(config.order), AddCombiner());
+  chain.Wait(true);
+  std::cerr << "=== 2/5 Calculating and sorting adjusted counts ===" << std::endl;
+  master.InitForAdjust(sorter, type_count);
+}
+
+void InitialProbabilities(const std::vector<uint64_t> &counts, const std::vector<Discount> &discounts, Master &master, Sorts<SuffixOrder> &primary, FixedArray<util::stream::FileBuffer> &gammas) {
+  const PipelineConfig &config = master.Config();
+  Chains second(config.order);
+
+  {
+    Sorts<ContextOrder> sorts;
+    master.SetupSorts(sorts);
+    PrintStatistics(counts, discounts);
+    lm::ngram::ShowSizes(counts);
+    std::cerr << "=== 3/5 Calculating and sorting initial probabilities ===" << std::endl;
+    master.SortAndReadTwice(counts, sorts, second, config.initial_probs.adder_in);
+  }
+
+  Chains gamma_chains(config.order);
+  InitialProbabilities(config.initial_probs, discounts, master.MutableChains(), second, gamma_chains);
+  // Don't care about gamma for 0.  
+  gamma_chains[0] >> util::stream::kRecycle;
+  gammas.Init(config.order - 1);
+  for (std::size_t i = 1; i < config.order; ++i) {
+    gammas.push_back(util::MakeTemp(config.TempPrefix()));
+    gamma_chains[i] >> gammas[i - 1].Sink();
+  }
+  // Has to be done here due to gamma_chains scope.
+  master.SetupSorts(primary);
+}
+
+void InterpolateProbabilities(const std::vector<uint64_t> &counts, Master &master, Sorts<SuffixOrder> &primary, FixedArray<util::stream::FileBuffer> &gammas) {
+  std::cerr << "=== 4/5 Calculating and writing order-interpolated probabilities ===" << std::endl;
+  const PipelineConfig &config = master.Config();
+  master.MaximumLazyInput(counts, primary);
+
+  Chains gamma_chains(config.order - 1);
+  util::stream::ChainConfig read_backoffs(config.read_backoffs);
+  read_backoffs.entry_size = sizeof(float);
+  for (std::size_t i = 0; i < config.order - 1; ++i) {
+    gamma_chains.push_back(read_backoffs);
+    gamma_chains.back() >> gammas[i].Source();
+  }
+  master >> Interpolate(counts[0], ChainPositions(gamma_chains));
+  gamma_chains >> util::stream::kRecycle;
+  master.BufferFinal(counts);
+}
+
+} // namespace
+
+void Pipeline(PipelineConfig config, int text_file, int out_arpa) {
+  // Some fail-fast sanity checks.
+  if (config.sort.buffer_size * 4 > config.TotalMemory()) {
+    config.sort.buffer_size = config.TotalMemory() / 4;
+    std::cerr << "Warning: changing sort block size to " << config.sort.buffer_size << " bytes due to low total memory." << std::endl;
+  }
+  if (config.minimum_block < NGram::TotalSize(config.order)) {
+    config.minimum_block = NGram::TotalSize(config.order);
+    std::cerr << "Warning: raising minimum block to " << config.minimum_block << " to fit an ngram in every block." << std::endl;
+  }
+  UTIL_THROW_IF(config.sort.buffer_size < config.minimum_block, util::Exception, "Sort block size " << config.sort.buffer_size << " is below the minimum block size " << config.minimum_block << ".");
+  UTIL_THROW_IF(config.TotalMemory() < config.minimum_block * config.order * config.block_count, util::Exception,
+      "Not enough memory to fit " << (config.order * config.block_count) << " blocks with minimum size " << config.minimum_block << ".  Increase memory to " << (config.minimum_block * config.order * config.block_count) << " bytes or decrease the minimum block size.");
+
+  UTIL_TIMER("(%w s) Total wall time elapsed\n");
+  Master master(config);
+
+  util::scoped_fd vocab_file(config.vocab_file.empty() ? 
+      util::MakeTemp(config.TempPrefix()) : 
+      util::CreateOrThrow(config.vocab_file.c_str()));
+  uint64_t token_count;
+  std::string text_file_name;
+  CountText(text_file, vocab_file.get(), master, token_count, text_file_name);
+
+  std::vector<uint64_t> counts;
+  std::vector<Discount> discounts;
+  master >> AdjustCounts(counts, discounts);
+
+  {
+    FixedArray<util::stream::FileBuffer> gammas;
+    Sorts<SuffixOrder> primary;
+    InitialProbabilities(counts, discounts, master, primary, gammas);
+    InterpolateProbabilities(counts, master, primary, gammas);
+  }
+
+  std::cerr << "=== 5/5 Writing ARPA model ===" << std::endl;
+  VocabReconstitute vocab(vocab_file.get());
+  UTIL_THROW_IF(vocab.Size() != counts[0], util::Exception, "Vocab words don't match up.  Is there a null byte in the input?");
+  HeaderInfo header_info(text_file_name, token_count);
+  master >> PrintARPA(vocab, counts, (config.verbose_header ? &header_info : NULL), out_arpa) >> util::stream::kRecycle;
+  master.MutableChains().Wait(true);
+}
+
+}} // namespaces
diff --git a/lm/builder/pipeline.hh b/lm/builder/pipeline.hh
new file mode 100644
index 000000000..f1d6c5f61
--- /dev/null
+++ b/lm/builder/pipeline.hh
@@ -0,0 +1,40 @@
+#ifndef LM_BUILDER_PIPELINE__
+#define LM_BUILDER_PIPELINE__
+
+#include "lm/builder/initial_probabilities.hh"
+#include "lm/builder/header_info.hh"
+#include "util/stream/config.hh"
+#include "util/file_piece.hh"
+
+#include <string>
+#include <cstddef>
+
+namespace lm { namespace builder {
+
+struct PipelineConfig {
+  std::size_t order;
+  std::string vocab_file;
+  util::stream::SortConfig sort;
+  InitialProbabilitiesConfig initial_probs;
+  util::stream::ChainConfig read_backoffs;
+  bool verbose_header;
+
+  // Amount of memory to assume that the vocabulary hash table will use.  This
+  // is subtracted from total memory for CorpusCount.
+  std::size_t assume_vocab_hash_size;
+
+  // Minimum block size to tolerate.
+  std::size_t minimum_block;
+
+  // Number of blocks to use.  This will be overridden to 1 if everything fits.
+  std::size_t block_count;
+
+  const std::string &TempPrefix() const { return sort.temp_prefix; }
+  std::size_t TotalMemory() const { return sort.total_memory; }
+};
+
+// Takes ownership of text_file.
+void Pipeline(PipelineConfig config, int text_file, int out_arpa);
+
+}} // namespaces
+#endif // LM_BUILDER_PIPELINE__
diff --git a/lm/builder/print.cc b/lm/builder/print.cc
new file mode 100644
index 000000000..b0323221a
--- /dev/null
+++ b/lm/builder/print.cc
@@ -0,0 +1,135 @@
+#include "lm/builder/print.hh"
+
+#include "util/double-conversion/double-conversion.h"
+#include "util/double-conversion/utils.h"
+#include "util/file.hh"
+#include "util/mmap.hh"
+#include "util/scoped.hh"
+#include "util/stream/timer.hh"
+
+#define BOOST_LEXICAL_CAST_ASSUME_C_LOCALE
+#include <boost/lexical_cast.hpp>
+
+#include <sstream>
+
+#include <string.h>
+
+namespace lm { namespace builder {
+
+VocabReconstitute::VocabReconstitute(int fd) {
+  uint64_t size = util::SizeOrThrow(fd);
+  util::MapRead(util::POPULATE_OR_READ, fd, 0, size, memory_);
+  const char *const start = static_cast<const char*>(memory_.get());
+  const char *i;
+  for (i = start; i != start + size; i += strlen(i) + 1) {
+    map_.push_back(i);
+  }
+  // Last one for LookupPiece.
+  map_.push_back(i);
+}
+
+namespace {
+class OutputManager {
+  public:
+    static const std::size_t kOutBuf = 1048576;
+
+    // Does not take ownership of out.
+    explicit OutputManager(int out)
+      : buf_(util::MallocOrThrow(kOutBuf)),
+        builder_(static_cast<char*>(buf_.get()), kOutBuf),
+        // Mostly the default but with inf instead.  And no flags.
+        convert_(double_conversion::DoubleToStringConverter::NO_FLAGS, "inf", "NaN", 'e', -6, 21, 6, 0),
+        fd_(out) {}
+
+    ~OutputManager() {
+      Flush();
+    }
+
+    OutputManager &operator<<(float value) {
+      // Odd, but this is the largest number found in the comments.
+      EnsureRemaining(double_conversion::DoubleToStringConverter::kMaxPrecisionDigits + 8);
+      convert_.ToShortestSingle(value, &builder_);
+      return *this;
+    }
+
+    OutputManager &operator<<(StringPiece str) {
+      if (str.size() > kOutBuf) {
+        Flush();
+        util::WriteOrThrow(fd_, str.data(), str.size());
+      } else {
+        EnsureRemaining(str.size());
+        builder_.AddSubstring(str.data(), str.size());
+      }
+      return *this;
+    }
+
+    // Inefficient!
+    OutputManager &operator<<(unsigned val) {
+      return *this << boost::lexical_cast<std::string>(val);
+    }
+
+    OutputManager &operator<<(char c) {
+      EnsureRemaining(1);
+      builder_.AddCharacter(c);
+      return *this;
+    }
+
+    void Flush() {
+      util::WriteOrThrow(fd_, buf_.get(), builder_.position());
+      builder_.Reset();
+    }
+
+  private:
+    void EnsureRemaining(std::size_t amount) {
+      if (static_cast<std::size_t>(builder_.size() - builder_.position()) < amount) {
+        Flush();
+      }
+    }
+
+    util::scoped_malloc buf_;
+    double_conversion::StringBuilder builder_;
+    double_conversion::DoubleToStringConverter convert_;
+    int fd_;
+};
+} // namespace
+
+PrintARPA::PrintARPA(const VocabReconstitute &vocab, const std::vector<uint64_t> &counts, const HeaderInfo* header_info, int out_fd) 
+  : vocab_(vocab), out_fd_(out_fd) {
+  std::stringstream stream;
+
+  if (header_info) {
+    stream << "# Input file: " << header_info->input_file << '\n';
+    stream << "# Token count: " << header_info->token_count << '\n';
+    stream << "# Smoothing: Modified Kneser-Ney" << '\n';
+  }
+  stream << "\\data\\\n";
+  for (size_t i = 0; i < counts.size(); ++i) {
+    stream << "ngram " << (i+1) << '=' << counts[i] << '\n';
+  }
+  stream << '\n';
+  std::string as_string(stream.str());
+  util::WriteOrThrow(out_fd, as_string.data(), as_string.size());
+}
+
+void PrintARPA::Run(const ChainPositions &positions) {
+  UTIL_TIMER("(%w s) Wrote ARPA file\n");
+  OutputManager out(out_fd_);
+  for (unsigned order = 1; order <= positions.size(); ++order) {
+    out << "\\" << order << "-grams:" << '\n';
+    for (NGramStream stream(positions[order - 1]); stream; ++stream) {
+      // Correcting for numerical precision issues.  Take that IRST.  
+      out << std::min(0.0f, stream->Value().complete.prob) << '\t' << vocab_.Lookup(*stream->begin());
+      for (const WordIndex *i = stream->begin() + 1; i != stream->end(); ++i) {
+        out << ' ' << vocab_.Lookup(*i);
+      }
+      float backoff = stream->Value().complete.backoff;
+      if (backoff != 0.0)
+        out << '\t' << backoff;
+      out << '\n';
+    }
+    out << '\n';
+  }
+  out << "\\end\\\n";
+}
+
+}} // namespaces
diff --git a/lm/builder/print.hh b/lm/builder/print.hh
new file mode 100644
index 000000000..aa932e757
--- /dev/null
+++ b/lm/builder/print.hh
@@ -0,0 +1,102 @@
+#ifndef LM_BUILDER_PRINT__
+#define LM_BUILDER_PRINT__
+
+#include "lm/builder/ngram.hh"
+#include "lm/builder/multi_stream.hh"
+#include "lm/builder/header_info.hh"
+#include "util/file.hh"
+#include "util/mmap.hh"
+#include "util/string_piece.hh"
+
+#include <ostream>
+
+#include <assert.h>
+
+// Warning: print routines read all unigrams before all bigrams before all
+// trigrams etc.  So if other parts of the chain move jointly, you'll have to
+// buffer.  
+
+namespace lm { namespace builder {
+
+class VocabReconstitute {
+  public:
+    // fd must be alive for life of this object; does not take ownership.
+    explicit VocabReconstitute(int fd);
+
+    const char *Lookup(WordIndex index) const {
+      assert(index < map_.size() - 1);
+      return map_[index];
+    }
+
+    StringPiece LookupPiece(WordIndex index) const {
+      return StringPiece(map_[index], map_[index + 1] - 1 - map_[index]);
+    }
+
+    std::size_t Size() const {
+      // There's an extra entry to support StringPiece lengths.
+      return map_.size() - 1;
+    }
+
+  private:
+    util::scoped_memory memory_;
+    std::vector<const char*> map_;
+};
+
+// Not defined, only specialized.  
+template <class T> void PrintPayload(std::ostream &to, const Payload &payload);
+template <> inline void PrintPayload<uint64_t>(std::ostream &to, const Payload &payload) {
+  to << payload.count;
+}
+template <> inline void PrintPayload<Uninterpolated>(std::ostream &to, const Payload &payload) {
+  to << log10(payload.uninterp.prob) << ' ' << log10(payload.uninterp.gamma);
+}
+template <> inline void PrintPayload<ProbBackoff>(std::ostream &to, const Payload &payload) {
+  to << payload.complete.prob << ' ' << payload.complete.backoff;
+}
+
+// template parameter is the type stored.  
+template <class V> class Print {
+  public:
+    explicit Print(const VocabReconstitute &vocab, std::ostream &to) : vocab_(vocab), to_(to) {}
+
+    void Run(const ChainPositions &chains) {
+      NGramStreams streams(chains);
+      for (NGramStream *s = streams.begin(); s != streams.end(); ++s) {
+        DumpStream(*s);
+      }
+    }
+
+    void Run(const util::stream::ChainPosition &position) {
+      NGramStream stream(position);
+      DumpStream(stream);
+    }
+
+  private:
+    void DumpStream(NGramStream &stream) {
+      for (; stream; ++stream) {
+        PrintPayload<V>(to_, stream->Value());
+        for (const WordIndex *w = stream->begin(); w != stream->end(); ++w) {
+          to_ << ' ' << vocab_.Lookup(*w) << '=' << *w;
+        }
+        to_ << '\n';
+      }
+    }
+
+    const VocabReconstitute &vocab_;
+    std::ostream &to_;
+};
+
+class PrintARPA {
+  public:
+    // header_info may be NULL to disable the header
+    explicit PrintARPA(const VocabReconstitute &vocab, const std::vector<uint64_t> &counts, const HeaderInfo* header_info, int out_fd);
+
+    void Run(const ChainPositions &positions);
+
+  private:
+    const VocabReconstitute &vocab_;
+    int out_fd_;
+};
+
+}} // namespaces
+#endif // LM_BUILDER_PRINT__
diff --git a/lm/builder/sort.hh b/lm/builder/sort.hh
new file mode 100644
index 000000000..9989389b6
--- /dev/null
+++ b/lm/builder/sort.hh
@@ -0,0 +1,103 @@
+#ifndef LM_BUILDER_SORT__
+#define LM_BUILDER_SORT__
+
+#include "lm/builder/multi_stream.hh"
+#include "lm/builder/ngram.hh"
+#include "lm/word_index.hh"
+#include "util/stream/sort.hh"
+
+#include "util/stream/timer.hh"
+
+#include <functional>
+#include <string>
+
+namespace lm {
+namespace builder {
+
+template <class Child> class Comparator : public std::binary_function<const void *, const void *, bool> {
+  public:
+    explicit Comparator(std::size_t order) : order_(order) {}
+
+    inline bool operator()(const void *lhs, const void *rhs) const {
+      return static_cast<const Child*>(this)->Compare(static_cast<const WordIndex*>(lhs), static_cast<const WordIndex*>(rhs));
+    }
+
+    std::size_t Order() const { return order_; }
+
+  protected:
+    std::size_t order_;
+};
+
+class SuffixOrder : public Comparator<SuffixOrder> {
+  public:
+    explicit SuffixOrder(std::size_t order) : Comparator<SuffixOrder>(order) {}
+
+    inline bool Compare(const WordIndex *lhs, const WordIndex *rhs) const {
+      for (std::size_t i = order_ - 1; i != 0; --i) {
+        if (lhs[i] != rhs[i])
+          return lhs[i] < rhs[i];
+      }
+      return lhs[0] < rhs[0];
+    }
+
+    static const unsigned kMatchOffset = 1;
+};
+
+class ContextOrder : public Comparator<ContextOrder> {
+  public:
+    explicit ContextOrder(std::size_t order) : Comparator<ContextOrder>(order) {}
+
+    inline bool Compare(const WordIndex *lhs, const WordIndex *rhs) const {
+      for (int i = order_ - 2; i >= 0; --i) {
+        if (lhs[i] != rhs[i])
+          return lhs[i] < rhs[i];
+      }
+      return lhs[order_ - 1] < rhs[order_ - 1];
+    }
+};
+
+class PrefixOrder : public Comparator<PrefixOrder> {
+  public:
+    explicit PrefixOrder(std::size_t order) : Comparator<PrefixOrder>(order) {}
+
+    inline bool Compare(const WordIndex *lhs, const WordIndex *rhs) const {
+      for (std::size_t i = 0; i < order_; ++i) {
+        if (lhs[i] != rhs[i])
+          return lhs[i] < rhs[i];
+      }
+      return false;
+    }
+    
+    static const unsigned kMatchOffset = 0;
+};
+
+// Sum counts for the same n-gram.
+struct AddCombiner {
+  bool operator()(void *first_void, const void *second_void, const SuffixOrder &compare) const {
+    NGram first(first_void, compare.Order());
+    // There isn't a const version of NGram.  
+    NGram second(const_cast<void*>(second_void), compare.Order());
+    if (memcmp(first.begin(), second.begin(), sizeof(WordIndex) * compare.Order())) return false;
+    first.Count() += second.Count();
+    return true;
+  }
+};
+
+// The combiner is only used on a single chain, so I didn't bother to allow
+// that template.  
+template <class Compare> class Sorts : public FixedArray<util::stream::Sort<Compare> > {
+  private:
+    typedef util::stream::Sort<Compare> S;
+    typedef FixedArray<S> P;
+
+  public:
+    void push_back(util::stream::Chain &chain, const util::stream::SortConfig &config, const Compare &compare) {
+      new (P::end()) S(chain, config, compare);
+      P::Constructed();
+    }
+};
+
+} // namespace builder
+} // namespace lm
+
+#endif // LM_BUILDER_SORT__
diff --git a/lm/filter/Jamfile b/lm/filter/Jamfile
new file mode 100644
index 000000000..16e417261
--- /dev/null
+++ b/lm/filter/Jamfile
@@ -0,0 +1,5 @@
+fakelib lm_filter : phrase.cc vocab.cc arpa_io.cc ../../util//kenutil : <threading>multi:<library>/top//boost_thread ;
+
+obj main : main.cc : <threading>single:<define>NTHREAD <include>../.. ;
+
+exe filter : main lm_filter ../../util//kenutil ..//kenlm : <threading>multi:<library>/top//boost_thread ;
diff --git a/lm/filter/arpa_io.cc b/lm/filter/arpa_io.cc
new file mode 100644
index 000000000..caf8df953
--- /dev/null
+++ b/lm/filter/arpa_io.cc
@@ -0,0 +1,122 @@
+#include "lm/filter/arpa_io.hh"
+#include "util/file_piece.hh"
+
+#include <iostream>
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include <ctype.h>
+#include <errno.h>
+#include <string.h>
+
+namespace lm {
+
+ARPAInputException::ARPAInputException(const StringPiece &message) throw() : what_("Error: ") {
+  what_.append(message.data(), message.size());
+}
+
+ARPAInputException::ARPAInputException(const StringPiece &message, const StringPiece &line) throw() {
+  what_ = "Error: ";
+  what_.append(message.data(), message.size());
+  what_ += " in line '";
+  what_.append(line.data(), line.size());
+  what_ += "'.";
+}
+
+ARPAOutputException::ARPAOutputException(const char *message, const std::string &file_name) throw()
+  : what_(std::string(message) + " file " + file_name), file_name_(file_name) {
+  if (errno) {
+    char buf[1024];
+    buf[0] = 0;
+#if (_POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600) && ! _GNU_SOURCE
+    const char *add = buf;
+    if (!strerror_r(errno, buf, 1024)) {
+#else
+    const char *add = strerror_r(errno, buf, 1024);
+    if (add) {
+#endif
+      what_ += " :";
+      what_ += add;
+    }
+  }
+}
+
+// Seeking is the responsibility of the caller.
+void WriteCounts(std::ostream &out, const std::vector<size_t> &number) {
+  out << "\n\\data\\\n";
+  for (unsigned int i = 0; i < number.size(); ++i) {
+    out << "ngram " << i+1 << "=" << number[i] << '\n';
+  }
+  out << '\n';
+}
+
+size_t SizeNeededForCounts(const std::vector<size_t> &number) {
+  std::ostringstream buf;
+  WriteCounts(buf, number);
+  return buf.tellp();
+}
+
+bool IsEntirelyWhiteSpace(const StringPiece &line) {
+  for (size_t i = 0; i < static_cast<size_t>(line.size()); ++i) {
+    if (!isspace(line.data()[i])) return false;
+  }
+  return true;
+}
+
+ARPAOutput::ARPAOutput(const char *name, size_t buffer_size) : file_name_(name), buffer_(new char[buffer_size]) {
+  try {
+    file_.exceptions(std::ostream::eofbit | std::ostream::failbit | std::ostream::badbit);
+    if (!file_.rdbuf()->pubsetbuf(buffer_.get(), buffer_size)) {
+      std::cerr << "Warning: could not enlarge buffer for " << name << std::endl;
+      buffer_.reset();
+    }
+    file_.open(name, std::ios::out | std::ios::binary);
+  } catch (const std::ios_base::failure &f) {
+    throw ARPAOutputException("Opening", file_name_);
+  }
+}
+
+void ARPAOutput::ReserveForCounts(std::streampos reserve) {
+  try {
+    for (std::streampos i = 0; i < reserve; i += std::streampos(1)) {
+      file_ << '\n';
+    }
+  } catch (const std::ios_base::failure &f) {
+    throw ARPAOutputException("Writing blanks to reserve space for counts to ", file_name_);
+  }
+}
+
+void ARPAOutput::BeginLength(unsigned int length) {
+  fast_counter_ = 0;
+  try {
+    file_ << '\\' << length << "-grams:" << '\n';
+  } catch (const std::ios_base::failure &f) {
+    throw ARPAOutputException("Writing n-gram header to ", file_name_);
+  }
+}
+
+void ARPAOutput::EndLength(unsigned int length) {
+  try {
+    file_ << '\n';
+  } catch (const std::ios_base::failure &f) {
+    throw ARPAOutputException("Writing blank at end of count list to ", file_name_);
+  }
+  if (length > counts_.size()) {
+    counts_.resize(length);
+  }
+  counts_[length - 1] = fast_counter_;
+}
+
+void ARPAOutput::Finish() {
+  try {
+    file_ << "\\end\\\n";
+    file_.seekp(0);
+    WriteCounts(file_, counts_);
+    file_ << std::flush;
+  } catch (const std::ios_base::failure &f) {
+    throw ARPAOutputException("Finishing including writing counts at beginning to ", file_name_);
+  }
+}
+
+} // namespace lm
diff --git a/lm/filter/arpa_io.hh b/lm/filter/arpa_io.hh
new file mode 100644
index 000000000..90f484473
--- /dev/null
+++ b/lm/filter/arpa_io.hh
@@ -0,0 +1,122 @@
+#ifndef LM_FILTER_ARPA_IO__
+#define LM_FILTER_ARPA_IO__
+/* Input and output for ARPA format language model files.
+ */
+#include "lm/read_arpa.hh"
+#include "util/exception.hh"
+#include "util/string_piece.hh"
+#include "util/tokenize_piece.hh"
+
+#include <boost/noncopyable.hpp>
+#include <boost/scoped_array.hpp>
+
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include <err.h>
+#include <string.h>
+
+namespace util { class FilePiece; }
+
+namespace lm {
+
+class ARPAInputException : public util::Exception {
+  public:
+    explicit ARPAInputException(const StringPiece &message) throw();
+    explicit ARPAInputException(const StringPiece &message, const StringPiece &line) throw();
+    virtual ~ARPAInputException() throw() {}
+
+    const char *what() const throw() { return what_.c_str(); }
+
+  private:
+    std::string what_;
+};
+
+class ARPAOutputException : public std::exception {
+  public:
+    ARPAOutputException(const char *prefix, const std::string &file_name) throw();
+    virtual ~ARPAOutputException() throw() {}
+
+    const char *what() const throw() { return what_.c_str(); }
+
+    const std::string &File() const throw() { return file_name_; }
+
+  private:
+    std::string what_;
+    const std::string file_name_;
+};
+
+// Handling for the counts of n-grams at the beginning of ARPA files.
+size_t SizeNeededForCounts(const std::vector<size_t> &number);
+
+/* Writes an ARPA file.  This has to be seekable so the counts can be written
+ * at the end.  Hence, I just have it own a std::fstream instead of accepting
+ * a separately held std::ostream.  
+ */
+class ARPAOutput : boost::noncopyable {
+  public:
+    explicit ARPAOutput(const char *name, size_t buffer_size = 65536);
+
+    void ReserveForCounts(std::streampos reserve);
+
+    void BeginLength(unsigned int length);
+
+    void AddNGram(const StringPiece &line) {
+      try {
+        file_ << line << '\n';
+      } catch (const std::ios_base::failure &f) {
+        throw ARPAOutputException("Writing an n-gram", file_name_);
+      }
+      ++fast_counter_;
+    }
+
+    void AddNGram(const StringPiece &ngram, const StringPiece &line) {
+      AddNGram(line);
+    }
+
+    template <class Iterator> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line) {
+      AddNGram(line);
+    }
+
+    void EndLength(unsigned int length);
+
+    void Finish();
+
+  private:
+    const std::string file_name_;
+    boost::scoped_array<char> buffer_;
+    std::fstream file_;
+    size_t fast_counter_;
+    std::vector<size_t> counts_;
+};
+
+
+template <class Output> void ReadNGrams(util::FilePiece &in, unsigned int length, size_t number, Output &out) {
+  ReadNGramHeader(in, length);
+  out.BeginLength(length);
+  for (size_t i = 0; i < number; ++i) {
+    StringPiece line = in.ReadLine();
+    util::TokenIter<util::SingleCharacter> tabber(line, '\t');
+    if (!tabber) throw ARPAInputException("blank line", line);
+    if (!++tabber) throw ARPAInputException("no tab", line);
+
+    out.AddNGram(*tabber, line);
+  }
+  out.EndLength(length);
+}
+
+template <class Output> void ReadARPA(util::FilePiece &in_lm, Output &out) {
+  std::vector<size_t> number;
+  ReadARPACounts(in_lm, number);
+  out.ReserveForCounts(SizeNeededForCounts(number));
+  for (unsigned int i = 0; i < number.size(); ++i) {
+    ReadNGrams(in_lm, i + 1, number[i], out);
+  }
+  ReadEnd(in_lm);
+  out.Finish();
+}
+
+} // namespace lm
+
+#endif // LM_FILTER_ARPA_IO__
diff --git a/lm/filter/count_io.hh b/lm/filter/count_io.hh
new file mode 100644
index 000000000..97c0fa25e
--- /dev/null
+++ b/lm/filter/count_io.hh
@@ -0,0 +1,91 @@
+#ifndef LM_FILTER_COUNT_IO__
+#define LM_FILTER_COUNT_IO__
+
+#include <fstream>
+#include <iostream>
+#include <string>
+
+#include <err.h>
+
+#include "util/file_piece.hh"
+
+namespace lm {
+
+class CountOutput : boost::noncopyable {
+  public:
+    explicit CountOutput(const char *name) : file_(name, std::ios::out) {}
+
+    void AddNGram(const StringPiece &line) {
+      if (!(file_ << line << '\n')) {
+        err(3, "Writing counts file failed");
+      }
+    }
+
+    template <class Iterator> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line) {
+      AddNGram(line);
+    }
+
+    void AddNGram(const StringPiece &ngram, const StringPiece &line) {
+      AddNGram(line);
+    }
+
+  private:
+    std::fstream file_;
+};
+
+class CountBatch {
+  public:
+    explicit CountBatch(std::streamsize initial_read) 
+      : initial_read_(initial_read) {
+      buffer_.reserve(initial_read);
+    }
+
+    void Read(std::istream &in) {
+      buffer_.resize(initial_read_);
+      in.read(&*buffer_.begin(), initial_read_);
+      buffer_.resize(in.gcount());
+      char got;
+      while (in.get(got) && got != '\n')
+        buffer_.push_back(got);
+    }
+
+    template <class Output> void Send(Output &out) {
+      for (util::TokenIter<util::SingleCharacter> line(StringPiece(&*buffer_.begin(), buffer_.size()), '\n'); line; ++line) {
+        util::TokenIter<util::SingleCharacter> tabber(*line, '\t');
+        if (!tabber) {
+          std::cerr << "Warning: empty n-gram count line being removed\n";
+          continue;
+        }
+        util::TokenIter<util::SingleCharacter, true> words(*tabber, ' ');
+        if (!words) {
+          std::cerr << "Line has a tab but no words.\n";
+          continue;
+        }
+        out.AddNGram(words, util::TokenIter<util::SingleCharacter, true>::end(), *line);
+      }
+    }
+
+  private:
+    std::streamsize initial_read_;
+
+    // This could have been a std::string but that's less happy with raw writes.  
+    std::vector<char> buffer_;
+};
+
+template <class Output> void ReadCount(util::FilePiece &in_file, Output &out) {
+  try {
+    while (true) {
+      StringPiece line = in_file.ReadLine();
+      util::TokenIter<util::SingleCharacter> tabber(line, '\t');
+      if (!tabber) {
+        std::cerr << "Warning: empty n-gram count line being removed\n";
+        continue;
+      }
+      out.AddNGram(*tabber, line);
+    }
+  } catch (const util::EndOfFileException &e) {}
+}
+
+} // namespace lm
+
+#endif // LM_FILTER_COUNT_IO__
diff --git a/lm/filter/format.hh b/lm/filter/format.hh
new file mode 100644
index 000000000..7f945b0d6
--- /dev/null
+++ b/lm/filter/format.hh
@@ -0,0 +1,250 @@
+#ifndef LM_FILTER_FORMAT_H__
+#define LM_FITLER_FORMAT_H__
+
+#include "lm/filter/arpa_io.hh"
+#include "lm/filter/count_io.hh"
+
+#include <boost/lexical_cast.hpp>
+#include <boost/ptr_container/ptr_vector.hpp>
+
+#include <iosfwd>
+
+namespace lm {
+
+template <class Single> class MultipleOutput {
+  private:
+    typedef boost::ptr_vector<Single> Singles;
+    typedef typename Singles::iterator SinglesIterator;
+
+  public:
+    MultipleOutput(const char *prefix, size_t number) {
+      files_.reserve(number);
+      std::string tmp;
+      for (unsigned int i = 0; i < number; ++i) {
+        tmp = prefix;
+        tmp += boost::lexical_cast<std::string>(i);
+        files_.push_back(new Single(tmp.c_str()));
+      }
+    }
+
+    void AddNGram(const StringPiece &line) {
+      for (SinglesIterator i = files_.begin(); i != files_.end(); ++i)
+        i->AddNGram(line);
+    }
+
+    template <class Iterator> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line) {
+      for (SinglesIterator i = files_.begin(); i != files_.end(); ++i)
+        i->AddNGram(begin, end, line);
+    }
+
+    void SingleAddNGram(size_t offset, const StringPiece &line) {
+      files_[offset].AddNGram(line);
+    }
+
+    template <class Iterator> void SingleAddNGram(size_t offset, const Iterator &begin, const Iterator &end, const StringPiece &line) {
+      files_[offset].AddNGram(begin, end, line);
+    }
+
+  protected:
+    Singles files_;
+};
+
+class MultipleARPAOutput : public MultipleOutput<ARPAOutput> {
+  public:
+    MultipleARPAOutput(const char *prefix, size_t number) : MultipleOutput<ARPAOutput>(prefix, number) {}
+
+    void ReserveForCounts(std::streampos reserve) {
+      for (boost::ptr_vector<ARPAOutput>::iterator i = files_.begin(); i != files_.end(); ++i)
+        i->ReserveForCounts(reserve);
+    }
+
+    void BeginLength(unsigned int length) {
+      for (boost::ptr_vector<ARPAOutput>::iterator i = files_.begin(); i != files_.end(); ++i)
+        i->BeginLength(length);
+    }
+
+    void EndLength(unsigned int length) {
+      for (boost::ptr_vector<ARPAOutput>::iterator i = files_.begin(); i != files_.end(); ++i)
+        i->EndLength(length);
+    }
+
+    void Finish() {
+      for (boost::ptr_vector<ARPAOutput>::iterator i = files_.begin(); i != files_.end(); ++i)
+        i->Finish();
+    }
+};
+
+template <class Filter, class Output> class DispatchInput {
+  public:
+    DispatchInput(Filter &filter, Output &output) : filter_(filter), output_(output) {}
+
+/*    template <class Iterator> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line) {
+      filter_.AddNGram(begin, end, line, output_);
+    }*/
+
+    void AddNGram(const StringPiece &ngram, const StringPiece &line) {
+      filter_.AddNGram(ngram, line, output_);
+    }
+
+  protected:
+    Filter &filter_;
+    Output &output_;
+};
+
+template <class Filter, class Output> class DispatchARPAInput : public DispatchInput<Filter, Output> {
+  private:
+    typedef DispatchInput<Filter, Output> B;
+
+  public:
+    DispatchARPAInput(Filter &filter, Output &output) : B(filter, output) {}
+
+    void ReserveForCounts(std::streampos reserve) { B::output_.ReserveForCounts(reserve); }
+    void BeginLength(unsigned int length) { B::output_.BeginLength(length); }
+
+    void EndLength(unsigned int length) {
+      B::filter_.Flush();
+      B::output_.EndLength(length);
+    }
+    void Finish() { B::output_.Finish(); }
+};
+
+struct ARPAFormat {
+  typedef ARPAOutput Output;
+  typedef MultipleARPAOutput Multiple;
+  static void Copy(util::FilePiece &in, Output &out) {
+    ReadARPA(in, out);
+  }
+  template <class Filter, class Out> static void RunFilter(util::FilePiece &in, Filter &filter, Out &output) {
+    DispatchARPAInput<Filter, Out> dispatcher(filter, output);
+    ReadARPA(in, dispatcher);
+  }
+};
+
+struct CountFormat {
+  typedef CountOutput Output;
+  typedef MultipleOutput<Output> Multiple;
+  static void Copy(util::FilePiece &in, Output &out) {
+    ReadCount(in, out);
+  }
+  template <class Filter, class Out> static void RunFilter(util::FilePiece &in, Filter &filter, Out &output) {
+    DispatchInput<Filter, Out> dispatcher(filter, output);
+    ReadCount(in, dispatcher);
+  }
+};
+
+/* For multithreading, the buffer classes hold batches of filter inputs and
+ * outputs in memory.  The strings get reused a lot, so keep them around
+ * instead of clearing each time.  
+ */
+class InputBuffer {
+  public:
+    InputBuffer() : actual_(0) {}
+    
+    void Reserve(size_t size) { lines_.reserve(size); }
+
+    template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
+      if (lines_.size() == actual_) lines_.resize(lines_.size() + 1);
+      // TODO avoid this copy.
+      std::string &copied = lines_[actual_].line;
+      copied.assign(line.data(), line.size());
+      lines_[actual_].ngram.set(copied.data() + (ngram.data() - line.data()), ngram.size());
+      ++actual_;
+    }
+
+    template <class Filter, class Output> void CallFilter(Filter &filter, Output &output) const {
+      for (std::vector<Line>::const_iterator i = lines_.begin(); i != lines_.begin() + actual_; ++i) {
+        filter.AddNGram(i->ngram, i->line, output);
+      }
+    }
+
+    void Clear() { actual_ = 0; }
+    bool Empty() { return actual_ == 0; }
+    size_t Size() { return actual_; }
+
+  private:
+    struct Line {
+      std::string line;
+      StringPiece ngram;
+    };
+
+    size_t actual_;
+
+    std::vector<Line> lines_;
+};
+
+class BinaryOutputBuffer {
+  public:
+    BinaryOutputBuffer() {}
+
+    void Reserve(size_t size) {
+      lines_.reserve(size);
+    }
+    
+    void AddNGram(const StringPiece &line) {
+      lines_.push_back(line);
+    }
+    
+    template <class Output> void Flush(Output &output) {
+      for (std::vector<StringPiece>::const_iterator i = lines_.begin(); i != lines_.end(); ++i) {
+        output.AddNGram(*i);
+      }
+      lines_.clear();
+    }
+    
+  private:
+    std::vector<StringPiece> lines_;
+};
+
+class MultipleOutputBuffer {
+  public:
+    MultipleOutputBuffer() : last_(NULL) {}
+
+    void Reserve(size_t size) {
+      annotated_.reserve(size);
+    }
+
+    void AddNGram(const StringPiece &line) {
+      annotated_.resize(annotated_.size() + 1);
+      annotated_.back().line = line;
+    }
+
+    void SingleAddNGram(size_t offset, const StringPiece &line) {
+      if ((line.data() == last_.data()) && (line.length() == last_.length())) {
+        annotated_.back().systems.push_back(offset);
+      } else {
+        annotated_.resize(annotated_.size() + 1);
+        annotated_.back().systems.push_back(offset);
+        annotated_.back().line = line;
+        last_ = line;
+      }
+    }
+
+    template <class Output> void Flush(Output &output) {
+      for (std::vector<Annotated>::const_iterator i = annotated_.begin(); i != annotated_.end(); ++i) {
+        if (i->systems.empty()) {
+          output.AddNGram(i->line);
+        } else {
+          for (std::vector<size_t>::const_iterator j = i->systems.begin(); j != i->systems.end(); ++j) {
+            output.SingleAddNGram(*j, i->line);
+          }
+        }
+      }
+      annotated_.clear();
+    }
+
+  private:
+    struct Annotated {
+      // If this is empty, send to all systems. 
+      // A filter should never send to all systems and send to a single one.
+      std::vector<size_t> systems;
+      StringPiece line;
+    };
+
+    StringPiece last_;
+
+    std::vector<Annotated> annotated_;
+};
+
+} // namespace lm
+
+#endif // LM_FILTER_FORMAT_H__
diff --git a/lm/filter/main.cc b/lm/filter/main.cc
new file mode 100644
index 000000000..c42243e2b
--- /dev/null
+++ b/lm/filter/main.cc
@@ -0,0 +1,249 @@
+#include "lm/filter/arpa_io.hh"
+#include "lm/filter/format.hh"
+#include "lm/filter/phrase.hh"
+#ifndef NTHREAD
+#include "lm/filter/thread.hh"
+#endif
+#include "lm/filter/vocab.hh"
+#include "lm/filter/wrapper.hh"
+#include "util/file_piece.hh"
+
+#include <boost/ptr_container/ptr_vector.hpp>
+
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <memory>
+
+namespace lm {
+namespace {
+
+void DisplayHelp(const char *name) {
+  std::cerr
+    << "Usage: " << name << " mode [context] [phrase] [raw|arpa] [threads:m] [batch_size:m] (vocab|model):input_file output_file\n\n"
+    "copy mode just copies, but makes the format nicer for e.g. irstlm's broken\n"
+    "    parser.\n"
+    "single mode treats the entire input as a single sentence.\n"
+    "multiple mode filters to multiple sentences in parallel.  Each sentence is on\n"
+    "    a separate line.  A separate file is created for each file by appending the\n"
+    "    0-indexed line number to the output file name.\n"
+    "union mode produces one filtered model that is the union of models created by\n"
+    "    multiple mode.\n\n"
+    "context means only the context (all but last word) has to pass the filter, but\n"
+    "    the entire n-gram is output.\n\n"
+    "phrase means that the vocabulary is actually tab-delimited phrases and that the\n"
+    "    phrases can generate the n-gram when assembled in arbitrary order and\n"
+    "    clipped.  Currently works with multiple or union mode.\n\n"
+    "The file format is set by [raw|arpa] with default arpa:\n"
+    "raw means space-separated tokens, optionally followed by a tab and arbitrary\n"
+    "    text.  This is useful for ngram count files.\n"
+    "arpa means the ARPA file format for n-gram language models.\n\n"
+#ifndef NTHREAD
+    "threads:m sets m threads (default: conccurrency detected by boost)\n"
+    "batch_size:m sets the batch size for threading.  Expect memory usage from this\n"
+    "    of 2*threads*batch_size n-grams.\n\n"
+#else
+    "This binary was compiled with -DNTHREAD, disabling threading.  If you wanted\n"
+    "    threading, compile without this flag against Boost >=1.42.0.\n\n"
+#endif
+    "There are two inputs: vocabulary and model.  Either may be given as a file\n"
+    "    while the other is on stdin.  Specify the type given as a file using\n"
+    "    vocab: or model: before the file name.  \n\n"
+    "For ARPA format, the output must be seekable.  For raw format, it can be a\n"
+    "    stream i.e. /dev/stdout\n";
+}
+
+typedef enum {MODE_COPY, MODE_SINGLE, MODE_MULTIPLE, MODE_UNION} FilterMode;
+typedef enum {FORMAT_ARPA, FORMAT_COUNT} Format;
+
+struct Config {
+  Config() : 
+#ifndef NTHREAD
+  batch_size(25000),
+  threads(boost::thread::hardware_concurrency()),
+#endif
+  phrase(false),
+  context(false),
+  format(FORMAT_ARPA)
+  {
+#ifndef NTHREAD
+    if (!threads) threads = 1;
+#endif
+  }
+
+#ifndef NTHREAD
+  size_t batch_size;
+  size_t threads;
+#endif
+  bool phrase;
+  bool context;
+  FilterMode mode;
+  Format format;
+};
+
+template <class Format, class Filter, class OutputBuffer, class Output> void RunThreadedFilter(const Config &config, util::FilePiece &in_lm, Filter &filter, Output &output) {
+#ifndef NTHREAD
+  if (config.threads == 1) {
+#endif
+    Format::RunFilter(in_lm, filter, output);
+#ifndef NTHREAD
+  } else {
+    typedef Controller<Filter, OutputBuffer, Output> Threaded;
+    Threaded threading(config.batch_size, config.threads * 2, config.threads, filter, output);
+    Format::RunFilter(in_lm, threading, output);
+  }
+#endif
+}
+
+template <class Format, class Filter, class OutputBuffer, class Output> void RunContextFilter(const Config &config, util::FilePiece &in_lm, Filter filter, Output &output) {
+  if (config.context) {
+    ContextFilter<Filter> context_filter(filter);
+    RunThreadedFilter<Format, ContextFilter<Filter>, OutputBuffer, Output>(config, in_lm, context_filter, output);
+  } else {
+    RunThreadedFilter<Format, Filter, OutputBuffer, Output>(config, in_lm, filter, output);
+  }
+}
+
+template <class Format, class Binary> void DispatchBinaryFilter(const Config &config, util::FilePiece &in_lm, const Binary &binary, typename Format::Output &out) {
+  typedef BinaryFilter<Binary> Filter;
+  RunContextFilter<Format, Filter, BinaryOutputBuffer, typename Format::Output>(config, in_lm, Filter(binary), out);
+}
+
+template <class Format> void DispatchFilterModes(const Config &config, std::istream &in_vocab, util::FilePiece &in_lm, const char *out_name) {
+  if (config.mode == MODE_MULTIPLE) {
+    if (config.phrase) {
+      typedef phrase::Multiple Filter;
+      phrase::Substrings substrings;
+      typename Format::Multiple out(out_name, phrase::ReadMultiple(in_vocab, substrings));
+      RunContextFilter<Format, Filter, MultipleOutputBuffer, typename Format::Multiple>(config, in_lm, Filter(substrings), out);
+    } else {
+      typedef vocab::Multiple Filter;
+      boost::unordered_map<std::string, std::vector<unsigned int> > words;
+      typename Format::Multiple out(out_name, vocab::ReadMultiple(in_vocab, words));
+      RunContextFilter<Format, Filter, MultipleOutputBuffer, typename Format::Multiple>(config, in_lm, Filter(words), out);
+    }
+    return;
+  }
+
+  typename Format::Output out(out_name);
+
+  if (config.mode == MODE_COPY) {
+    Format::Copy(in_lm, out);
+    return;
+  }
+
+  if (config.mode == MODE_SINGLE) {
+    vocab::Single::Words words;
+    vocab::ReadSingle(in_vocab, words);
+    DispatchBinaryFilter<Format, vocab::Single>(config, in_lm, vocab::Single(words), out);
+    return;
+  }
+
+  if (config.mode == MODE_UNION) {
+    if (config.phrase) {
+      phrase::Substrings substrings;
+      phrase::ReadMultiple(in_vocab, substrings);
+      DispatchBinaryFilter<Format, phrase::Union>(config, in_lm, phrase::Union(substrings), out);
+    } else {
+      vocab::Union::Words words;
+      vocab::ReadMultiple(in_vocab, words);
+      DispatchBinaryFilter<Format, vocab::Union>(config, in_lm, vocab::Union(words), out);
+    }
+    return;
+  }
+}
+
+} // namespace
+} // namespace lm
+
+int main(int argc, char *argv[]) {
+  if (argc < 4) {
+    lm::DisplayHelp(argv[0]);
+    return 1;
+  }
+
+  // I used to have boost::program_options, but some users didn't want to compile boost.  
+  lm::Config config;
+  boost::optional<lm::FilterMode> mode;
+  for (int i = 1; i < argc - 2; ++i) {
+    const char *str = argv[i];
+    if (!std::strcmp(str, "copy")) {
+      mode = lm::MODE_COPY;
+    } else if (!std::strcmp(str, "single")) {
+      mode = lm::MODE_SINGLE;
+    } else if (!std::strcmp(str, "multiple")) {
+      mode = lm::MODE_MULTIPLE;
+    } else if (!std::strcmp(str, "union")) {
+      mode = lm::MODE_UNION;
+    } else if (!std::strcmp(str, "phrase")) {
+      config.phrase = true;
+    } else if (!std::strcmp(str, "context")) {
+      config.context = true;
+    } else if (!std::strcmp(str, "arpa")) {
+      config.format = lm::FORMAT_ARPA;
+    } else if (!std::strcmp(str, "raw")) {
+      config.format = lm::FORMAT_COUNT;
+#ifndef NTHREAD
+    } else if (!std::strncmp(str, "threads:", 8)) {
+      config.threads = boost::lexical_cast<size_t>(str + 8);
+      if (!config.threads) {
+        std::cerr << "Specify at least one thread." << std::endl;
+        return 1;
+      }
+    } else if (!std::strncmp(str, "batch_size:", 11)) {
+      config.batch_size = boost::lexical_cast<size_t>(str + 11);
+      if (config.batch_size < 5000) {
+        std::cerr << "Batch size must be at least one and should probably be >= 5000" << std::endl;
+        if (!config.batch_size) return 1;
+      }
+#endif
+    } else {
+      lm::DisplayHelp(argv[0]);
+      return 1;
+    }
+  }
+  
+  if (!mode) {
+    lm::DisplayHelp(argv[0]);
+    return 1;
+  }
+  config.mode = *mode;
+
+  if (config.phrase && config.mode != lm::MODE_UNION && mode != lm::MODE_MULTIPLE) {
+    std::cerr << "Phrase constraint currently only works in multiple or union mode.  If you really need it for single, put everything on one line and use union." << std::endl;
+    return 1;
+  }
+
+  bool cmd_is_model = true;
+  const char *cmd_input = argv[argc - 2];
+  if (!strncmp(cmd_input, "vocab:", 6)) {
+    cmd_is_model = false;
+    cmd_input += 6;
+  } else if (!strncmp(cmd_input, "model:", 6)) {
+    cmd_input += 6;
+  } else if (strchr(cmd_input, ':')) {
+    errx(1, "Specify vocab: or model: before the input file name, not \"%s\"", cmd_input);
+  } else {
+    std::cerr << "Assuming that " << cmd_input << " is a model file" << std::endl;
+  }
+  std::ifstream cmd_file;
+  std::istream *vocab;
+  if (cmd_is_model) {
+    vocab = &std::cin;
+  } else {
+    cmd_file.open(cmd_input, std::ios::in);
+    if (!cmd_file) {
+      err(2, "Could not open input file %s", cmd_input);
+    }
+    vocab = &cmd_file;
+  }
+
+  util::FilePiece model(cmd_is_model ? util::OpenReadOrThrow(cmd_input) : 0, cmd_is_model ? cmd_input : NULL, &std::cerr);
+
+  if (config.format == lm::FORMAT_ARPA) {
+    lm::DispatchFilterModes<lm::ARPAFormat>(config, *vocab, model, argv[argc - 1]);
+  } else if (config.format == lm::FORMAT_COUNT) {
+    lm::DispatchFilterModes<lm::CountFormat>(config, *vocab, model, argv[argc - 1]);
+  }
+  return 0;
+}
diff --git a/lm/filter/phrase.cc b/lm/filter/phrase.cc
new file mode 100644
index 000000000..1bef2a3f3
--- /dev/null
+++ b/lm/filter/phrase.cc
@@ -0,0 +1,281 @@
+#include "lm/filter/phrase.hh"
+
+#include "lm/filter/format.hh"
+
+#include <algorithm>
+#include <functional>
+#include <iostream>
+#include <queue>
+#include <string>
+#include <vector>
+
+#include <ctype.h>
+
+namespace lm {
+namespace phrase {
+
+unsigned int ReadMultiple(std::istream &in, Substrings &out) {
+  bool sentence_content = false;
+  unsigned int sentence_id = 0;
+  std::vector<Hash> phrase;
+  std::string word;
+  while (in) {
+    char c;
+    // Gather a word.
+    while (!isspace(c = in.get()) && in) word += c;
+    // Treat EOF like a newline.
+    if (!in) c = '\n';
+    // Add the word to the phrase.
+    if (!word.empty()) {
+      phrase.push_back(util::MurmurHashNative(word.data(), word.size()));
+      word.clear();
+    }
+    if (c == ' ') continue;
+    // It's more than just a space.  Close out the phrase.  
+    if (!phrase.empty()) {
+      sentence_content = true;
+      out.AddPhrase(sentence_id, phrase.begin(), phrase.end());
+      phrase.clear();
+    }
+    if (c == '\t' || c == '\v') continue;
+    // It's more than a space or tab: a newline.   
+    if (sentence_content) {
+      ++sentence_id;
+      sentence_content = false;
+    }
+  }
+  if (!in.eof()) in.exceptions(std::istream::failbit | std::istream::badbit);
+  return sentence_id + sentence_content;
+}
+
+namespace detail { const StringPiece kEndSentence("</s>"); }
+
+namespace {
+
+typedef unsigned int Sentence;
+typedef std::vector<Sentence> Sentences;
+
+class Vertex;
+
+class Arc {
+  public:
+    Arc() {}
+
+    // For arcs from one vertex to another.  
+    void SetPhrase(Vertex &from, Vertex &to, const Sentences &intersect) {
+      Set(to, intersect);
+      from_ = &from;
+    }
+
+    /* For arcs from before the n-gram begins to somewhere in the n-gram (right
+     * aligned).  These have no from_ vertex; it implictly matches every
+     * sentence.  This also handles when the n-gram is a substring of a phrase. 
+     */
+    void SetRight(Vertex &to, const Sentences &complete) {
+      Set(to, complete);
+      from_ = NULL;
+    }
+
+    Sentence Current() const {
+      return *current_;
+    }
+
+    bool Empty() const {
+      return current_ == last_;
+    }
+
+    /* When this function returns:
+     * If Empty() then there's nothing left from this intersection.
+     *
+     * If Current() == to then to is part of the intersection. 
+     *
+     * Otherwise, Current() > to.  In this case, to is not part of the
+     * intersection and neither is anything < Current().  To determine if
+     * any value >= Current() is in the intersection, call LowerBound again
+     * with the value.   
+     */
+    void LowerBound(const Sentence to);
+
+  private:
+    void Set(Vertex &to, const Sentences &sentences);
+
+    const Sentence *current_;
+    const Sentence *last_;
+    Vertex *from_;
+};
+
+struct ArcGreater : public std::binary_function<const Arc *, const Arc *, bool> {
+  bool operator()(const Arc *first, const Arc *second) const {
+    return first->Current() > second->Current();
+  }
+};
+
+class Vertex {
+  public:
+    Vertex() : current_(0) {}
+
+    Sentence Current() const {
+      return current_;
+    }
+
+    bool Empty() const {
+      return incoming_.empty();
+    }
+
+    void LowerBound(const Sentence to);
+
+  private:
+    friend class Arc;
+
+    void AddIncoming(Arc *arc) {
+      if (!arc->Empty()) incoming_.push(arc);
+    }
+
+    unsigned int current_;
+    std::priority_queue<Arc*, std::vector<Arc*>, ArcGreater> incoming_;
+};
+
+void Arc::LowerBound(const Sentence to) {
+  current_ = std::lower_bound(current_, last_, to);
+  // If *current_ > to, don't advance from_.  The intervening values of
+  // from_ may be useful for another one of its outgoing arcs.
+  if (!from_ || Empty() || (Current() > to)) return;
+  assert(Current() == to);
+  from_->LowerBound(to);
+  if (from_->Empty()) {
+    current_ = last_;
+    return;
+  }
+  assert(from_->Current() >= to);
+  if (from_->Current() > to) {
+    current_ = std::lower_bound(current_ + 1, last_, from_->Current());
+  }
+}
+
+void Arc::Set(Vertex &to, const Sentences &sentences) {
+  current_ = &*sentences.begin();
+  last_ = &*sentences.end();
+  to.AddIncoming(this);
+}
+
+void Vertex::LowerBound(const Sentence to) {
+  if (Empty()) return;
+  // Union lower bound.  
+  while (true) {
+    Arc *top = incoming_.top();
+    if (top->Current() > to) {
+      current_ = top->Current();
+      return;
+    }
+    // If top->Current() == to, we still need to verify that's an actual 
+    // element and not just a bound.  
+    incoming_.pop();
+    top->LowerBound(to);
+    if (!top->Empty()) {
+      incoming_.push(top);
+      if (top->Current() == to) {
+        current_ = to;
+        return;
+      }
+    } else if (Empty()) {
+      return;
+    }
+  }
+}
+
+void BuildGraph(const Substrings &phrase, const std::vector<Hash> &hashes, Vertex *const vertices, Arc *free_arc) {
+  assert(!hashes.empty());
+
+  const Hash *const first_word = &*hashes.begin();
+  const Hash *const last_word = &*hashes.end() - 1;
+
+  Hash hash = 0;
+  const Sentences *found;
+  // Phrases starting at or before the first word in the n-gram.
+  {
+    Vertex *vertex = vertices;
+    for (const Hash *word = first_word; ; ++word, ++vertex) {
+      hash = util::MurmurHashNative(&hash, sizeof(uint64_t), *word);
+      // Now hash is [hashes.begin(), word].
+      if (word == last_word) {
+        if (phrase.FindSubstring(hash, found))
+          (free_arc++)->SetRight(*vertex, *found);
+        break;
+      }
+      if (!phrase.FindRight(hash, found)) break;
+      (free_arc++)->SetRight(*vertex, *found);
+    }
+  }
+
+  // Phrases starting at the second or later word in the n-gram.   
+  Vertex *vertex_from = vertices;
+  for (const Hash *word_from = first_word + 1; word_from != &*hashes.end(); ++word_from, ++vertex_from) {
+    hash = 0;
+    Vertex *vertex_to = vertex_from + 1;
+    for (const Hash *word_to = word_from; ; ++word_to, ++vertex_to) {
+      // Notice that word_to and vertex_to have the same index.  
+      hash = util::MurmurHashNative(&hash, sizeof(uint64_t), *word_to);
+      // Now hash covers [word_from, word_to].
+      if (word_to == last_word) {
+        if (phrase.FindLeft(hash, found))
+          (free_arc++)->SetPhrase(*vertex_from, *vertex_to, *found);
+        break;
+      }
+      if (!phrase.FindPhrase(hash, found)) break;
+      (free_arc++)->SetPhrase(*vertex_from, *vertex_to, *found);
+    }
+  }
+}
+
+} // namespace
+
+namespace detail {
+
+} // namespace detail
+
+bool Union::Evaluate() {
+  assert(!hashes_.empty());
+  // Usually there are at most 6 words in an n-gram, so stack allocation is reasonable.  
+  Vertex vertices[hashes_.size()];
+  // One for every substring.  
+  Arc arcs[((hashes_.size() + 1) * hashes_.size()) / 2];
+  BuildGraph(substrings_, hashes_, vertices, arcs);
+  Vertex &last_vertex = vertices[hashes_.size() - 1];
+
+  unsigned int lower = 0;
+  while (true) {
+    last_vertex.LowerBound(lower);
+    if (last_vertex.Empty()) return false;
+    if (last_vertex.Current() == lower) return true;
+    lower = last_vertex.Current();
+  }
+}
+
+template <class Output> void Multiple::Evaluate(const StringPiece &line, Output &output) {
+  assert(!hashes_.empty());
+  // Usually there are at most 6 words in an n-gram, so stack allocation is reasonable.  
+  Vertex vertices[hashes_.size()];
+  // One for every substring.  
+  Arc arcs[((hashes_.size() + 1) * hashes_.size()) / 2];
+  BuildGraph(substrings_, hashes_, vertices, arcs);
+  Vertex &last_vertex = vertices[hashes_.size() - 1];
+
+  unsigned int lower = 0;
+  while (true) {
+    last_vertex.LowerBound(lower);
+    if (last_vertex.Empty()) return;
+    if (last_vertex.Current() == lower) {
+      output.SingleAddNGram(lower, line);
+      ++lower;
+    } else {
+      lower = last_vertex.Current();
+    }
+  }
+}
+
+template void Multiple::Evaluate<CountFormat::Multiple>(const StringPiece &line, CountFormat::Multiple &output);
+template void Multiple::Evaluate<ARPAFormat::Multiple>(const StringPiece &line, ARPAFormat::Multiple &output);
+template void Multiple::Evaluate<MultipleOutputBuffer>(const StringPiece &line, MultipleOutputBuffer &output);
+
+} // namespace phrase
+} // namespace lm
diff --git a/lm/filter/phrase.hh b/lm/filter/phrase.hh
new file mode 100644
index 000000000..07479dea6
--- /dev/null
+++ b/lm/filter/phrase.hh
@@ -0,0 +1,153 @@
+#ifndef LM_FILTER_PHRASE_H__
+#define LM_FILTER_PHRASE_H__
+
+#include "util/murmur_hash.hh"
+#include "util/string_piece.hh"
+#include "util/tokenize_piece.hh"
+
+#include <boost/unordered_map.hpp>
+
+#include <iosfwd>
+#include <vector>
+
+#define LM_FILTER_PHRASE_METHOD(caps, lower) \
+bool Find##caps(Hash key, const std::vector<unsigned int> *&out) const {\
+  Table::const_iterator i(table_.find(key));\
+  if (i==table_.end()) return false; \
+  out = &i->second.lower; \
+  return true; \
+}
+
+namespace lm {
+namespace phrase {
+
+typedef uint64_t Hash;
+
+class Substrings {
+  private:
+    /* This is the value in a hash table where the key is a string.  It indicates
+     * four sets of sentences:
+     * substring is sentences with a phrase containing the key as a substring.  
+     * left is sentencess with a phrase that begins with the key (left aligned).
+     * right is sentences with a phrase that ends with the key (right aligned).
+     * phrase is sentences where the key is a phrase.
+     * Each set is encoded as a vector of sentence ids in increasing order.
+     */
+    struct SentenceRelation {
+      std::vector<unsigned int> substring, left, right, phrase;
+    };
+    /* Most of the CPU is hash table lookups, so let's not complicate it with
+     * vector equality comparisons.  If a collision happens, the SentenceRelation
+     * structure will contain the union of sentence ids over the colliding strings.
+     * In that case, the filter will be slightly more permissive.  
+     * The key here is the same as boost's hash of std::vector<std::string>.  
+     */
+    typedef boost::unordered_map<Hash, SentenceRelation> Table;
+
+  public:
+    Substrings() {}
+
+    /* If the string isn't a substring of any phrase, return NULL.  Otherwise,
+     * return a pointer to std::vector<unsigned int> listing sentences with
+     * matching phrases.  This set may be empty for Left, Right, or Phrase.
+     * Example: const std::vector<unsigned int> *FindSubstring(Hash key)
+     */
+    LM_FILTER_PHRASE_METHOD(Substring, substring)
+    LM_FILTER_PHRASE_METHOD(Left, left)
+    LM_FILTER_PHRASE_METHOD(Right, right)
+    LM_FILTER_PHRASE_METHOD(Phrase, phrase)
+
+    // sentence_id must be non-decreasing.  Iterators are over words in the phrase.  
+    template <class Iterator> void AddPhrase(unsigned int sentence_id, const Iterator &begin, const Iterator &end) {
+      // Iterate over all substrings.  
+      for (Iterator start = begin; start != end; ++start) {
+        Hash hash = 0;
+        SentenceRelation *relation;
+        for (Iterator finish = start; finish != end; ++finish) {
+          hash = util::MurmurHashNative(&hash, sizeof(uint64_t), *finish);
+          // Now hash is of [start, finish].
+          relation = &table_[hash];
+          AppendSentence(relation->substring, sentence_id);
+          if (start == begin) AppendSentence(relation->left, sentence_id);
+        }
+        AppendSentence(relation->right, sentence_id);
+        if (start == begin) AppendSentence(relation->phrase, sentence_id);
+      }
+    }
+
+  private:
+    void AppendSentence(std::vector<unsigned int> &vec, unsigned int sentence_id) {
+      if (vec.empty() || vec.back() != sentence_id) vec.push_back(sentence_id);
+    }
+
+    Table table_;
+};
+
+// Read a file with one sentence per line containing tab-delimited phrases of
+// space-separated words.  
+unsigned int ReadMultiple(std::istream &in, Substrings &out);
+
+namespace detail {
+extern const StringPiece kEndSentence;
+
+template <class Iterator> void MakeHashes(Iterator i, const Iterator &end, std::vector<Hash> &hashes) {
+  hashes.clear();
+  if (i == end) return;
+  // TODO: check strict phrase boundaries after <s> and before </s>.  For now, just skip tags.  
+  if ((i->data()[0] == '<') && (i->data()[i->size() - 1] == '>')) {
+    ++i;
+  }
+  for (; i != end && (*i != kEndSentence); ++i) {
+    hashes.push_back(util::MurmurHashNative(i->data(), i->size()));
+  }
+}
+
+} // namespace detail
+
+class Union {
+  public:
+    explicit Union(const Substrings &substrings) : substrings_(substrings) {}
+
+    template <class Iterator> bool PassNGram(const Iterator &begin, const Iterator &end) {
+      detail::MakeHashes(begin, end, hashes_);
+      return hashes_.empty() || Evaluate();
+    }
+
+  private:
+    bool Evaluate();
+
+    std::vector<Hash> hashes_;
+
+    const Substrings &substrings_;
+};
+
+class Multiple {
+  public:
+    explicit Multiple(const Substrings &substrings) : substrings_(substrings) {}
+
+    template <class Iterator, class Output> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line, Output &output) {
+      detail::MakeHashes(begin, end, hashes_);
+      if (hashes_.empty()) {
+        output.AddNGram(line);
+        return;
+      }
+      Evaluate(line, output);
+    }
+
+    template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
+      AddNGram(util::TokenIter<util::SingleCharacter, true>(ngram, ' '), util::TokenIter<util::SingleCharacter, true>::end(), line, output);
+    }
+
+    void Flush() const {}
+
+  private:
+    template <class Output> void Evaluate(const StringPiece &line, Output &output);
+
+    std::vector<Hash> hashes_;
+
+    const Substrings &substrings_;
+};
+
+} // namespace phrase
+} // namespace lm
+#endif // LM_FILTER_PHRASE_H__
diff --git a/lm/filter/thread.hh b/lm/filter/thread.hh
new file mode 100644
index 000000000..e785b2637
--- /dev/null
+++ b/lm/filter/thread.hh
@@ -0,0 +1,167 @@
+#ifndef LM_FILTER_THREAD_H__
+#define LM_FILTER_THREAD_H__
+
+#include "util/thread_pool.hh"
+
+#include <boost/utility/in_place_factory.hpp>
+
+#include <deque>
+#include <stack>
+
+namespace lm {
+
+template <class OutputBuffer> class ThreadBatch {
+  public:
+    ThreadBatch() {}
+    
+    void Reserve(size_t size) {
+      input_.Reserve(size);
+      output_.Reserve(size);
+     }
+
+    // File reading thread.  
+    InputBuffer &Fill(uint64_t sequence) {
+      sequence_ = sequence;
+      // Why wait until now to clear instead of after output?  free in the same
+      // thread as allocated.  
+      input_.Clear();
+      return input_;
+    }
+
+    // Filter worker thread.  
+    template <class Filter> void CallFilter(Filter &filter) {
+      input_.CallFilter(filter, output_);
+    }
+
+    uint64_t Sequence() const { return sequence_; }
+
+    // File writing thread.  
+    template <class RealOutput> void Flush(RealOutput &output) {
+      output_.Flush(output);
+    }
+
+  private:
+    InputBuffer input_;
+    OutputBuffer output_;
+
+    uint64_t sequence_;
+};
+
+template <class Batch, class Filter> class FilterWorker {
+  public:
+    typedef Batch *Request;
+
+    FilterWorker(const Filter &filter, util::PCQueue<Request> &done) : filter_(filter), done_(done) {}
+
+    void operator()(Request request) {
+      request->CallFilter(filter_);
+      done_.Produce(request);
+    }
+
+  private:
+    Filter filter_;
+
+    util::PCQueue<Request> &done_;
+};
+
+// There should only be one OutputWorker.
+template <class Batch, class Output> class OutputWorker {
+  public:
+    typedef Batch *Request;
+
+    OutputWorker(Output &output, util::PCQueue<Request> &done) : output_(output), done_(done), base_sequence_(0) {}
+
+    void operator()(Request request) {
+      assert(request->Sequence() >= base_sequence_);
+      // Assemble the output in order.  
+      uint64_t pos = request->Sequence() - base_sequence_;
+      if (pos >= ordering_.size()) {
+        ordering_.resize(pos + 1, NULL);
+      }
+      ordering_[pos] = request;
+      while (!ordering_.empty() && ordering_.front()) {
+        ordering_.front()->Flush(output_);
+        done_.Produce(ordering_.front());
+        ordering_.pop_front();
+        ++base_sequence_;
+      }
+    }
+
+  private:
+    Output &output_;
+
+    util::PCQueue<Request> &done_;
+
+    std::deque<Request> ordering_;
+
+    uint64_t base_sequence_;
+};
+
+template <class Filter, class OutputBuffer, class RealOutput> class Controller : boost::noncopyable {
+  private:
+    typedef ThreadBatch<OutputBuffer> Batch;
+
+  public:
+    Controller(size_t batch_size, size_t queue, size_t workers, const Filter &filter, RealOutput &output) 
+      : batch_size_(batch_size), queue_size_(queue),
+        batches_(queue),
+        to_read_(queue),
+        output_(queue, 1, boost::in_place(boost::ref(output), boost::ref(to_read_)), NULL),
+        filter_(queue, workers, boost::in_place(boost::ref(filter), boost::ref(output_.In())), NULL),
+        sequence_(0) {
+      for (size_t i = 0; i < queue; ++i) {
+        batches_[i].Reserve(batch_size);
+        local_read_.push(&batches_[i]);
+      }
+      NewInput();
+    }
+
+    void AddNGram(const StringPiece &ngram, const StringPiece &line, RealOutput &output) {
+      input_->AddNGram(ngram, line, output);
+      if (input_->Size() == batch_size_) {
+        FlushInput();
+        NewInput();
+      }
+    }
+
+    void Flush() {
+      FlushInput();
+      while (local_read_.size() < queue_size_) {
+        MoveRead();
+      }
+      NewInput();
+    }
+
+  private:
+    void FlushInput() {
+      if (input_->Empty()) return;
+      filter_.Produce(local_read_.top());
+      local_read_.pop();
+      if (local_read_.empty()) MoveRead();
+    }
+
+    void NewInput() {
+      input_ = &local_read_.top()->Fill(sequence_++);
+    }
+
+    void MoveRead() {
+      local_read_.push(to_read_.Consume());
+    }
+
+    const size_t batch_size_;
+    const size_t queue_size_;
+
+    std::vector<Batch> batches_;
+
+    util::PCQueue<Batch*> to_read_;
+    std::stack<Batch*> local_read_;
+    util::ThreadPool<OutputWorker<Batch, RealOutput> > output_;
+    util::ThreadPool<FilterWorker<Batch, Filter> > filter_;
+
+    uint64_t sequence_;
+    InputBuffer *input_;
+};
+
+} // namespace lm
+
+#endif // LM_FILTER_THREAD_H__
diff --git a/lm/filter/vocab.cc b/lm/filter/vocab.cc
new file mode 100644
index 000000000..7ee4e84ba
--- /dev/null
+++ b/lm/filter/vocab.cc
@@ -0,0 +1,54 @@
+#include "lm/filter/vocab.hh"
+
+#include <istream>
+#include <iostream>
+
+#include <ctype.h>
+#include <err.h>
+
+namespace lm {
+namespace vocab {
+
+void ReadSingle(std::istream &in, boost::unordered_set<std::string> &out) {
+  in.exceptions(std::istream::badbit);
+  std::string word;
+  while (in >> word) {
+    out.insert(word);
+  }
+}
+
+namespace {
+bool IsLineEnd(std::istream &in) {
+  int got;
+  do {
+    got = in.get();
+    if (!in) return true;
+    if (got == '\n') return true;
+  } while (isspace(got));
+  in.unget();
+  return false;
+}
+}// namespace
+
+// Read space separated words in enter separated lines.  These lines can be
+// very long, so don't read an entire line at a time.  
+unsigned int ReadMultiple(std::istream &in, boost::unordered_map<std::string, std::vector<unsigned int> > &out) {
+  in.exceptions(std::istream::badbit);
+  unsigned int sentence = 0;
+  bool used_id = false;
+  std::string word;
+  while (in >> word) {
+    used_id = true;
+    std::vector<unsigned int> &posting = out[word];
+    if (posting.empty() || (posting.back() != sentence))
+      posting.push_back(sentence);
+    if (IsLineEnd(in)) {
+      ++sentence;
+      used_id = false;
+    }
+  }
+  return sentence + used_id;
+}
+
+} // namespace vocab
+} // namespace lm
diff --git a/lm/filter/vocab.hh b/lm/filter/vocab.hh
new file mode 100644
index 000000000..e2b6adffa
--- /dev/null
+++ b/lm/filter/vocab.hh
@@ -0,0 +1,132 @@
+#ifndef LM_FILTER_VOCAB_H__
+#define LM_FILTER_VOCAB_H__
+
+// Vocabulary-based filters for language models.
+
+#include "util/multi_intersection.hh"
+#include "util/string_piece.hh"
+#include "util/tokenize_piece.hh"
+
+#include <boost/noncopyable.hpp>
+#include <boost/range/iterator_range.hpp>
+#include <boost/unordered/unordered_map.hpp>
+#include <boost/unordered/unordered_set.hpp>
+
+#include <string>
+#include <vector>
+
+namespace lm {
+namespace vocab {
+
+void ReadSingle(std::istream &in, boost::unordered_set<std::string> &out);
+
+// Read one sentence vocabulary per line.  Return the number of sentences.
+unsigned int ReadMultiple(std::istream &in, boost::unordered_map<std::string, std::vector<unsigned int> > &out);
+
+/* Is this a special tag like <s> or <UNK>?  This actually includes anything
+ * surrounded with < and >, which most tokenizers separate for real words, so
+ * this should not catch real words as it looks at a single token.   
+ */
+inline bool IsTag(const StringPiece &value) {
+  // The parser should never give an empty string.
+  assert(!value.empty());
+  return (value.data()[0] == '<' && value.data()[value.size() - 1] == '>');
+}
+
+class Single {
+  public:
+    typedef boost::unordered_set<std::string> Words;
+
+    explicit Single(const Words &vocab) : vocab_(vocab) {}
+
+    template <class Iterator> bool PassNGram(const Iterator &begin, const Iterator &end) {
+      for (Iterator i = begin; i != end; ++i) {
+        if (IsTag(*i)) continue;
+        if (FindStringPiece(vocab_, *i) == vocab_.end()) return false;
+      }
+      return true;
+    }
+
+  private:
+    const Words &vocab_;
+};
+
+class Union {
+  public:
+    typedef boost::unordered_map<std::string, std::vector<unsigned int> > Words;
+
+    explicit Union(const Words &vocabs) : vocabs_(vocabs) {}
+
+    template <class Iterator> bool PassNGram(const Iterator &begin, const Iterator &end) {
+      sets_.clear();
+
+      for (Iterator i(begin); i != end; ++i) {
+        if (IsTag(*i)) continue;
+        Words::const_iterator found(FindStringPiece(vocabs_, *i));
+        if (vocabs_.end() == found) return false;
+        sets_.push_back(boost::iterator_range<const unsigned int*>(&*found->second.begin(), &*found->second.end()));
+      }
+      return (sets_.empty() || util::FirstIntersection(sets_));
+    }
+
+  private:
+    const Words &vocabs_;
+
+    std::vector<boost::iterator_range<const unsigned int*> > sets_;
+};
+
+class Multiple {
+  public:
+    typedef boost::unordered_map<std::string, std::vector<unsigned int> > Words;
+
+    Multiple(const Words &vocabs) : vocabs_(vocabs) {}
+
+  private:
+    // Callback from AllIntersection that does AddNGram.
+    template <class Output> class Callback {
+      public:
+        Callback(Output &out, const StringPiece &line) : out_(out), line_(line) {}
+
+        void operator()(unsigned int index) {
+          out_.SingleAddNGram(index, line_);
+        }
+
+      private:
+        Output &out_;
+        const StringPiece &line_;
+    };
+
+  public:
+    template <class Iterator, class Output> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line, Output &output) {
+      sets_.clear();
+      for (Iterator i(begin); i != end; ++i) {
+        if (IsTag(*i)) continue;
+        Words::const_iterator found(FindStringPiece(vocabs_, *i));
+        if (vocabs_.end() == found) return;
+        sets_.push_back(boost::iterator_range<const unsigned int*>(&*found->second.begin(), &*found->second.end()));
+      }
+      if (sets_.empty()) {
+        output.AddNGram(line);
+        return;
+      }
+
+      Callback<Output> cb(output, line);
+      util::AllIntersection(sets_, cb);
+    }
+
+    template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
+      AddNGram(util::TokenIter<util::SingleCharacter, true>(ngram, ' '), util::TokenIter<util::SingleCharacter, true>::end(), line, output);
+    }
+
+    void Flush() const {}
+
+  private:
+    const Words &vocabs_;
+
+    std::vector<boost::iterator_range<const unsigned int*> > sets_;
+};
+
+} // namespace vocab
+} // namespace lm
+
+#endif // LM_FILTER_VOCAB_H__
diff --git a/lm/filter/wrapper.hh b/lm/filter/wrapper.hh
new file mode 100644
index 000000000..90b07a08f
--- /dev/null
+++ b/lm/filter/wrapper.hh
@@ -0,0 +1,58 @@
+#ifndef LM_FILTER_WRAPPER_H__
+#define LM_FILTER_WRAPPER_H__
+
+#include "util/string_piece.hh"
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+namespace lm {
+
+// Provide a single-output filter with the same interface as a
+// multiple-output filter so clients code against one interface.
+template <class Binary> class BinaryFilter {
+  public:
+    // Binary modes are just references (and a set) and it makes the API cleaner to copy them.  
+    explicit BinaryFilter(Binary binary) : binary_(binary) {}
+
+    template <class Iterator, class Output> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line, Output &output) {
+      if (binary_.PassNGram(begin, end))
+        output.AddNGram(line);
+    }
+
+    template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
+      AddNGram(util::TokenIter<util::SingleCharacter, true>(ngram, ' '), util::TokenIter<util::SingleCharacter, true>::end(), line, output);
+    }
+
+    void Flush() const {}
+
+  private:
+    Binary binary_;
+};
+
+// Wrap another filter to pay attention only to context words
+template <class FilterT> class ContextFilter {
+  public:
+    typedef FilterT Filter;
+
+    explicit ContextFilter(Filter &backend) : backend_(backend) {}
+
+    template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
+      pieces_.clear();
+      // TODO: this copy could be avoided by a lookahead iterator.
+      std::copy(util::TokenIter<util::SingleCharacter, true>(ngram, ' '), util::TokenIter<util::SingleCharacter, true>::end(), std::back_insert_iterator<std::vector<StringPiece> >(pieces_));
+      backend_.AddNGram(pieces_.begin(), pieces_.end() - !pieces_.empty(), line, output);
+    }
+
+    void Flush() const {}
+
+  private:
+    std::vector<StringPiece> pieces_;
+
+    Filter backend_;
+};
+
+} // namespace lm
+
+#endif // LM_FILTER_WRAPPER_H__
diff --git a/lm/read_arpa.cc b/lm/read_arpa.cc
index 4723ab3a2..9ea087985 100644
--- a/lm/read_arpa.cc
+++ b/lm/read_arpa.cc
@@ -46,8 +46,14 @@ uint64_t ReadCount(const std::string &from) {
 
 void ReadARPACounts(util::FilePiece &in, std::vector<uint64_t> &number) {
   number.clear();
-  StringPiece line;
-  while (IsEntirelyWhiteSpace(line = in.ReadLine())) {}
+  StringPiece line = in.ReadLine();
+  // In general, ARPA files can have arbitrary text before "\data\"
+  // But in KenLM, we require such lines to start with "#", so that
+  // we can do stricter error checking
+  while (IsEntirelyWhiteSpace(line) || line.starts_with("#")) {
+    line = in.ReadLine();
+  }
+
   if (line != "\\data\\") {
     if ((line.size() >= 2) && (line.data()[0] == 0x1f) && (static_cast<unsigned char>(line.data()[1]) == 0x8b)) {
       UTIL_THROW(FormatLoadException, "Looks like a gzip file.  If this is an ARPA file, pipe " << in.FileName() << " through zcat.  If this already in binary format, you need to decompress it because mmap doesn't work on top of gzip.");
diff --git a/lm/sizes.cc b/lm/sizes.cc
new file mode 100644
index 000000000..55ad586c4
--- /dev/null
+++ b/lm/sizes.cc
@@ -0,0 +1,63 @@
+#include "lm/sizes.hh"
+#include "lm/model.hh"
+#include "util/file_piece.hh"
+
+#include <vector>
+#include <iomanip>
+
+namespace lm {
+namespace ngram {
+
+void ShowSizes(const std::vector<uint64_t> &counts, const lm::ngram::Config &config) {
+  uint64_t sizes[6];
+  sizes[0] = ProbingModel::Size(counts, config);
+  sizes[1] = RestProbingModel::Size(counts, config);
+  sizes[2] = TrieModel::Size(counts, config);
+  sizes[3] = QuantTrieModel::Size(counts, config);
+  sizes[4] = ArrayTrieModel::Size(counts, config);
+  sizes[5] = QuantArrayTrieModel::Size(counts, config);
+  uint64_t max_length = *std::max_element(sizes, sizes + sizeof(sizes) / sizeof(uint64_t));
+  uint64_t min_length = *std::min_element(sizes, sizes + sizeof(sizes) / sizeof(uint64_t));
+  uint64_t divide;
+  char prefix;
+  if (min_length < (1 << 10) * 10) {
+    prefix = ' ';
+    divide = 1;
+  } else if (min_length < (1 << 20) * 10) {
+    prefix = 'k';
+    divide = 1 << 10;
+  } else if (min_length < (1ULL << 30) * 10) {
+    prefix = 'M';
+    divide = 1 << 20;
+  } else {
+    prefix = 'G';
+    divide = 1 << 30;
+  }
+  long int length = std::max<long int>(2, static_cast<long int>(ceil(log10((double) max_length / divide))));
+  std::cerr << "Memory estimate for binary LM:\ntype    ";
+
+  // right align bytes.  
+  for (long int i = 0; i < length - 2; ++i) std::cerr << ' ';
+
+  std::cerr << prefix << "B\n"
+    "probing " << std::setw(length) << (sizes[0] / divide) << " assuming -p " << config.probing_multiplier << "\n"
+    "probing " << std::setw(length) << (sizes[1] / divide) << " assuming -r models -p " << config.probing_multiplier << "\n"
+    "trie    " << std::setw(length) << (sizes[2] / divide) << " without quantization\n"
+    "trie    " << std::setw(length) << (sizes[3] / divide) << " assuming -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits << " quantization \n"
+    "trie    " << std::setw(length) << (sizes[4] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " array pointer compression\n"
+    "trie    " << std::setw(length) << (sizes[5] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits<< " array pointer compression and quantization\n";
+}
+
+void ShowSizes(const std::vector<uint64_t> &counts) {
+  lm::ngram::Config config;
+  ShowSizes(counts, config);
+}
+
+void ShowSizes(const char *file, const lm::ngram::Config &config) {
+  std::vector<uint64_t> counts;
+  util::FilePiece f(file);
+  lm::ReadARPACounts(f, counts);
+  ShowSizes(counts, config);
+}
+
+}} //namespaces
diff --git a/lm/sizes.hh b/lm/sizes.hh
new file mode 100644
index 000000000..85abade77
--- /dev/null
+++ b/lm/sizes.hh
@@ -0,0 +1,17 @@
+#ifndef LM_SIZES__
+#define LM_SIZES__
+
+#include <vector>
+
+#include <stdint.h>
+
+namespace lm { namespace ngram {
+
+struct Config;
+
+void ShowSizes(const std::vector<uint64_t> &counts, const lm::ngram::Config &config);
+void ShowSizes(const std::vector<uint64_t> &counts);
+void ShowSizes(const char *file, const lm::ngram::Config &config);
+
+}} // namespaces
+#endif // LM_SIZES__