1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
|
#!/usr/bin/env ruby
# frozen_string_literal: true
require 'gitlab'
require 'json'
class Reporter
IDENTIFIABLE_NOTE_TAG = 'gitlab-org/ai-powered/ai-framework:duo-chat-qa-evaluation-'
GRADE_TO_EMOJI_MAPPING = {
correct: ":white_check_mark:",
incorrect: ":x:",
unexpected: ":warning:"
}.freeze
def run
merge_request_iid = ENV['CI_MERGE_REQUEST_IID']
ci_project_id = ENV['CI_PROJECT_ID']
puts "Saving #{artifact_path}"
File.write(artifact_path, report_note)
# Look for an existing note
report_notes = com_gitlab_client
.merge_request_notes(ci_project_id, merge_request_iid)
.auto_paginate
.select do |note|
note.body.include? note_identifier_tag
end
note = report_notes.max_by { |note| Time.parse(note.created_at) }
if note && note.type != 'DiscussionNote'
# The latest note has not led to a discussion. Update it.
com_gitlab_client.edit_merge_request_note(ci_project_id, merge_request_iid, note.id, report_note)
puts "Updated comment."
else
# This is the first note or the latest note has been discussed on the MR.
# Don't update, create new note instead.
com_gitlab_client.create_merge_request_note(ci_project_id, merge_request_iid, report_note)
puts "Posted comment."
end
end
private
def report_filename
"#{ENV['DUO_RSPEC']}.md"
end
def artifact_path
File.join(ENV['CI_PROJECT_DIR'], report_filename)
end
def note_identifier_tag
"#{IDENTIFIABLE_NOTE_TAG}#{ENV['DUO_RSPEC']}"
end
def com_gitlab_client
@com_gitlab_client ||= Gitlab.client(
endpoint: "https://gitlab.com/api/v4",
private_token: ENV['PROJECT_TOKEN_FOR_CI_SCRIPTS_API_USAGE']
)
end
def report_note
report = <<~MARKDOWN
<!-- #{note_identifier_tag} -->
## GitLab Duo Chat QA evaluation
Report generated for "#{ENV['CI_JOB_NAME']}". This report is generated and refreshed automatically. Do not edit.
LLMs have been asked to evaluate GitLab Duo Chat's answers.
:white_check_mark: : LLM evaluated the answer as `CORRECT`.
:x: : LLM evaluated the answer as `INCORRECT`.
:warning: : LLM did not evaluate correctly or the evaluation request might have failed.
### Summary
- The total number of evaluations: #{summary_numbers[:total]}
- The number of evaluations in which all LLMs graded `CORRECT`: #{summary_numbers[:correct]} (#{summary_numbers[:correct_ratio]}%)
- Note: if an evaluation request failed or its response was not parsable, it was ignored. For example, :white_check_mark: :warning: would count as `CORRECT`.
- The number of evaluations in which all LLMs graded `INCORRECT`: #{summary_numbers[:incorrect]} (#{summary_numbers[:incorrect_ratio]}%)
- Note: if an evaluation request failed or its response was not parsable, it was ignored. For example, :x: :warning: would count as `INCORRECT`.
- The number of evaluations in which LLMs disagreed: #{summary_numbers[:disagreed]} (#{summary_numbers[:disagreed_ratio]}%)
### Evaluations
#{eval_content}
MARKDOWN
if report.length > 1000000
return <<~MARKDOWN
<!-- #{note_identifier_tag} -->
## GitLab Duo Chat QA evaluation
Report generated for "#{ENV['CI_JOB_NAME']}". This report is generated and refreshed automatically. Do not edit.
**:warning: the evaluation report is too long (> `1000000`) and cannot be posted as a note.**
Please check out the artifact for the CI job "#{ENV['CI_JOB_NAME']}":
https://gitlab.com/gitlab-org/gitlab/-/jobs/#{ENV['CI_JOB_ID']}/artifacts/file/#{report_filename}
MARKDOWN
end
report
end
def report_data
@report_data ||= Dir[File.join(ENV['CI_PROJECT_DIR'], "tmp/duo_chat/qa*.json")]
.map { |file| JSON.parse(File.read(file)) }
end
def eval_content
report_data
.sort_by { |a| a["question"] }
.map do |data|
<<~MARKDOWN
<details>
<summary>
#{correctness_indicator(data)}
`"#{data['question']}"`
(context: `#{data['resource']}`)
</summary>
#### Resource
`#{data['resource']}`
#### Answer
#{data['answer']}
#### LLM Evaluation
#{evalutions(data)}
</details>
MARKDOWN
end
.join
end
def summary_numbers
@graded_evaluations ||= report_data.map { |data| data["evaluations"].map { |eval| parse_grade(eval) } }
total = @graded_evaluations.size
correct = @graded_evaluations.count { |grades| !(grades.include? :incorrect) }
incorrect = @graded_evaluations.count { |grades| !(grades.include? :correct) }
disagreed = @graded_evaluations.count { |grades| (grades.include? :correct) && (grades.include? :incorrect) }
{
total: total,
correct: correct,
correct_ratio: (correct.to_f / total * 100).round(1),
incorrect: incorrect,
incorrect_ratio: (incorrect.to_f / total * 100).round(1),
disagreed: disagreed,
disagreed_ratio: (disagreed.to_f / total * 100).round(1)
}
end
def parse_grade(eval)
return :correct if eval["response"].match?(/Grade: CORRECT/i)
return :incorrect if eval["response"].match?(/Grade: INCORRECT/i)
# If the LLM's evaluation includes neither CORRECT nor CORRECT, flag it.
:unexpected
end
def correctness_indicator(data)
data["evaluations"].map { |eval| parse_grade(eval) }.map { |grade| GRADE_TO_EMOJI_MAPPING[grade] }.join(' ')
end
def evalutions(data)
rows = data["evaluations"].map do |eval|
grade = parse_grade(eval)
<<~MARKDOWN
<tr>
<td>#{eval['model']}</td>
<td>
#{GRADE_TO_EMOJI_MAPPING[grade]}
</td>
<td>
#{eval['response']}
</td
</tr>
MARKDOWN
end
.join
<<~MARKDOWN
<table>
<tr>
<td>Model</td>
<td>Grade</td>
<td>Details</td>
</tr>
#{rows}
</table>
MARKDOWN
end
end
Reporter.new.run
|