Skip to content

docs(evals): correct the temperature guidance in the config schema #207

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
adnanrhussain merged 1 commit into from
Aug 27, 2026
Merged

docs(evals): correct the temperature guidance in the config schema #207

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
2 changes: 1 addition & 1 deletion evals/_schemas/config.schema.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -155,7 +155,7 @@
"type": ["number", "null"],
"minimum": 0,
"maximum": 1,
"description": "Sampling temperature to send, or null to send nothing so the model uses its own default. Use null for models that require it: Gemini 3 (Google recommends removing the parameter; below 1.0 risks looping and degraded reasoning), GPT-5 (rejects any value but 1 with a 400), and Claude Opus 5 / Opus 4.8 / Opus 4.7 / Sonnet 5 / Fable 5 (sampling parameters removed, 400). Models that still accept one -- Gemini 2.x, GPT-4.x, Claude 4.6 and below -- should carry an explicit number. Required rather than optional so every step records a deliberate choice; an absent key reads as 'nobody considered it'."
"description": "Sampling temperature to send, or null to omit the parameter entirely. Prefer an explicit number: the value an evaluator was validated at should be recorded, not inherited from a model default that can change under us. Use null only for models that reject any non-default value -- Claude Opus 4.7 and later return a 400 -- where omitting is the documented approach. A model recommending its default is not grounds for null; write the number out. Required rather than optional so every step records a deliberate choice; an absent key reads as 'nobody considered it'."
}
}
},
Expand Down