Skip to content

docs: explain how to aggregate inference results - #17

Closed
Liang-Chun Tsai (ltsai-dev) wants to merge 2 commits into
mainfrom
ltsai-microsoft-docs-analyzing-results-guide
Closed

docs: explain how to aggregate inference results#17
Liang-Chun Tsai (ltsai-dev) wants to merge 2 commits into
mainfrom
ltsai-microsoft-docs-analyzing-results-guide

Conversation

@ltsai-dev

@ltsai-dev Liang-Chun Tsai (ltsai-dev) commented Aug 17, 2026

Copy link
Copy Markdown
Contributor

Summary

  • add a concise guide for aggregating tb infer JSONL results
  • document the single tb agg command used to obtain pass@1, pass@20, and pass^20
  • link the guide from the README documentation index

The guide intentionally refers readers to the main README for generating the JSONL and omits unrelated schema, filtering, debugging, and failure-analysis details.

Liang-Chun Tsai and others added 2 commits August 17, 2026 15:52
Documents the tb infer JSONL artifact schema, tb agg/tb pp/tb run-test/tb sbs
commands, and the pass@1/pass@20/pass^20 metrics aligned with the ThinkingBox
paper terminology, including the pass^k naming caveat (biased plug-in vs
true unbiased estimator). Covers domain/split filtering via jq/python on the
JSONL output, resuming interrupted runs via --previous-results-file, and
explicitly distinguishes built-in CLI evidence-gathering from the paper's
manual/external four-category failure classification.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: f71fc0b5-25dc-43ae-8eea-04d0e45f8de9
@ltsai-dev Liang-Chun Tsai (ltsai-dev) changed the title docs: add analyzing_results.md guide for tb infer / tb agg output-and-analysis workflow docs: explain how to aggregate inference results Aug 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant