Skip to content

Studio: stop the model from replying twice when it refuses #5775

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

Merged
danielhanchen merged 1 commit into from
May 26, 2026
+3 −1
Merged

Studio: stop the model from replying twice when it refuses #5775

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
4 changes: 3 additions & 1 deletion studio/backend/core/inference/llama_cpp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,9 @@
# Handles both straight and curly apostrophes.
# Excludes "I can", "I should", "I want to", "let's" which
# appear frequently in direct answers / explanations.
r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b"
# Negative lookahead drops negated forms ("I will not", "I'll never")
# so a refusal doesn't trigger a re-prompt.
r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
Comment on lines +63 to +65

@gemini-code-assist gemini-code-assist Bot May 26, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

While the negative lookahead for not and never successfully handles the most common refusal patterns, models can also refuse using other common formal phrasings such as "decline", "refuse", or "be unable".

Expanding the negative lookahead to cover these variations makes the refusal detection more robust and prevents duplicate responses when the model refuses using these alternative phrasings.

Suggested change
# Negative lookahead drops negated forms ("I will not", "I'll never")
# so a refusal doesn't trigger a re-prompt.
r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
# Negative lookahead drops negated or refusal forms ("I will not", "I'll never", "I will decline", "I will be unable")
# so a refusal doesn't trigger a re-prompt.
r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never|decline|declining|refuse|refusing|be\s+unable)\b)"

@oobabooga oobabooga May 26, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A refusal-verb list is bypassed by one intervening word ("I will respectfully decline", "I will have to decline"), so it never converges. The general case belongs in the "don't re-prompt after a final answer" path (the #5714 direction), not in this regex. Keeping it to "not"/"never", which handles the common case.

r"|"
# Step/plan framing: "First ...", "Step 1:", "Here's my plan"
r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
Expand Down