Datasets:

---

compressionawareintelligence

/

cai-bench

Like 10

Makes running CAI-Bench and sharing results easier now that the dataset is gated.

• caibench_share.py (new): packages a run_api.py / run_local.py run into a submission folder (submission.json, raw/, scores.json). Sends nothing unless --open-pr is passed and the user answers y; then opens a PR under the user's own HF account. Replaces a custom base_url with "custom", hashes data files against SHA256SUMS, refuses payloads that look like they contain an API key. Standard library only (except --open-pr).
• run_api.py, run_local.py: if data/ is missing, explain the gated download (form, login, hf download) instead of a FileNotFoundError; print the share hint at the end of a run.
• score_headline.py: print the share hint after the scores (not with --json).
• README.md: Access note, five-command quick start, "no frontier model on the leaderboard yet" call to action, login note in Load it, working share row.
• results/README.md: caibench_share.py fast path; raw/ allowed instead of replies.jsonl; no contact field (PRs are public). Published results section unchanged.