Align BrowseComp and Odysseys evaluators#67
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2cfeda9a82
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| return ( | ||
| f"User Task (context only): {task}\n\n" | ||
| "Evaluate ONLY this rubric item:\n" | ||
| + "\n".join(rubric_lines) | ||
| + f"\n\nFull Action History:\n{action_history}\n\n" | ||
| f"Screenshots attached below: {screenshot_count} " |
There was a problem hiding this comment.
This new per-rubric prompt no longer includes the agent's final answer even though evaluate_one still passes it into grade_rubrics. Many Odysseys rubrics require verifying what the final response states or returns, so the grader can mark those rubrics from screenshots/actions alone and miss cases where the browser reached the right page but the submitted answer omitted or misstated the required fields.
Useful? React with 👍 / 👎.
| else: | ||
| self.last_usage = None | ||
| return "correct: no" | ||
| raise |
There was a problem hiding this comment.
When a BrowseComp grading request exhausts retries, this re-raises an OpenAI error; BaseEvaluator._run_iteration catches openai.OpenAIError and continues without appending an EvalResult. In runs with a transient, credential, or rate-limit failure for a task, that task disappears from the results and summary denominator instead of being recorded as failed, which can make BrowseComp accuracy look better than the attempted evaluation.
Useful? React with 👍 / 👎.
1 participant
Summary
Test