Skip to content

fix(llm_judge): use last-match for verdict extraction in single and pairwise judging #3921

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
AUTHENSOR wants to merge 1 commit into
base: main
Choose a base branch
from
+16 −16
Open

fix(llm_judge): use last-match for verdict extraction in single and pairwise judging #3921

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
32 changes: 16 additions & 16 deletions fastchat/llm_judge/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,12 +173,12 @@ def run_judge_single(question, answer, judge, ref_answer, multi_turn=False):
raise ValueError(f"Invalid judge model name: {model}")

if judge.prompt_template["output_format"] == "[[rating]]":
match = re.search(one_score_pattern, judgment)
if not match:
match = re.search(one_score_pattern_backup, judgment)
matches = re.findall(one_score_pattern, judgment)
if not matches:
matches = re.findall(one_score_pattern_backup, judgment)

if match:
rating = ast.literal_eval(match.groups()[0])
if matches:
rating = ast.literal_eval(matches[-1])
else:
rating = -1
else:
Expand Down Expand Up @@ -280,20 +280,20 @@ def run_judge_pair(question, answer_a, answer_b, judge, ref_answer, multi_turn=F
raise ValueError(f"Invalid judge model name: {model}")

if judge.prompt_template["output_format"] == "[[A]]":
if "[[A]]" in judgment:
winner = "A"
elif "[[B]]" in judgment:
winner = "B"
elif "[[C]]" in judgment:
winner = "tie"
# Use findall and take the last match so that a verdict token echoed
# in reasoning does not override the judge's final verdict.
pairwise_matches = re.findall(r"\[\[([ABC])\]\]", judgment.upper())
if pairwise_matches:
verdict = pairwise_matches[-1]
winner = {"A": "A", "B": "B", "C": "tie"}.get(verdict, "error")
else:
winner = "error"
elif judge.prompt_template["output_format"] == "[[rating_a,rating_b]]":
match = re.search(two_score_pattern, judgment)
if not match:
match = re.search(two_score_pattern_backup, judgment)
if match:
scores = [ast.literal_eval(s.strip()) for s in match.groups()]
matches = re.findall(two_score_pattern, judgment)
if not matches:
matches = re.findall(two_score_pattern_backup, judgment)
if matches:
scores = [ast.literal_eval(s.strip()) for s in matches[-1]]
if abs(scores[0] - scores[1]) <= TIE_DELTA:
winner = "tie"
elif scores[0] > scores[1]:
Expand Down
Loading