Skip to content

judge: clamp max_tokens to the model's output cap (gpt-4o-mini 400s on 32768) #13

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
saikrishnarallabandi wants to merge 1 commit into
base: main
Choose a base branch
from
Open

judge: clamp max_tokens to the model's output cap (gpt-4o-mini 400s on 32768) #13

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
8 changes: 6 additions & 2 deletions src/claw_anything/graders/llm_judge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -152,7 +152,11 @@ def _chat_with_retries(
{"role": "user", "content": user},
],
"temperature": 0.0,
"max_tokens": max_tokens,
# Clamp to the model's real output cap. The default 32768 is
# above gpt-4o-mini's 16384 limit; ollama silently ignores an
# oversized max_tokens but OpenAI 400s, and the 20x retry loop
# then hangs grading until the batch drops the result.
"max_tokens": min(max_tokens, 16384),
}
resp = self.client.chat.completions.create(**kwargs)
log_llm_call(
Expand Down Expand Up @@ -334,7 +338,7 @@ def evaluate(
{"role": "user", "content": user_msg},
],
"temperature": 0.0,
"max_tokens": 32768,
"max_tokens": 16384, # gpt-4o-mini output cap; see the note above
}
resp = self.client.chat.completions.create(**_judge_kwargs)
log_llm_call(
Expand Down