Evaluation Report Skill for ChatGPT: Compare a Baseline and Flag Regressions
Aggregate supplied evaluation results by task or model. Show comparable scores, denominators, and threshold misses, then identify what to retest.
中文Install in ChatGPT
Use an eligible ChatGPT workspace with permission to create skills and, for uploads, upload permission. The installation path for this text-only adaptation was checked against OpenAI documentation. Workspace runtime verification is pending.
- Download agentic/code/addons/aiwg-evals/skills/eval-report/SKILL.md from the pinned GitHub source. Keep the name, description, and instructions for eval-report together as a standalone text skill.
- In a ChatGPT workspace with skills and uploads enabled, open Plugins, Skills, Create, and Upload from your computer. Choose a skill package format accepted by the file picker. You can also choose Create with editor and enter the original name, description, and instructions.
- After the scan completes, check that eval-report appears in Installed. Type @ in a new chat, select the skill, and use the fictional example on this page to check the output.
Example task
A fictional assistant answered the same evaluation tasks. Summary improved from 8/10 to 9/10. Format fell from 9/10 to 7/10. Flag regressions with an 80% threshold.What to check
Report a 10-point gain for summaries and a 20-point drop below threshold for format. Preserve denominators and conditions, and avoid calling ten cases a stable trend.
Original SKILL.md on GitHub · Updated
Frequently asked question
Can I compare the new version if its tasks changed?
Report the sets separately and note the change. Repeat both versions on the same questions before interpreting the difference as comparable performance.