PAWBench-Results

PAWBench Eval Package 20260708

This dataset package defines the clean HF repository layout and contains the public metadata needed to run the PAWBench Full31 coverage and Table1 PAW-Cal PAWEval rerun inputs. In the local no-copy handoff, videos are not duplicated under this directory; they are uploaded into sharded videos/by_media_id/ subdirectories from the private upload plan after explicit human approval.

Layout

• videos/by_media_id/<sha-prefix>/<media_id>.mp4: one physical video per media_id after human upload from the private upload plan.
• manifests/shared_media_registry.jsonl: public media identity registry.
• manifests/full31_paweval_manifest.jsonl: Full31 logical PAWEval rows.
• manifests/table1_paweval_rerun_rows.jsonl: Table1 PAW-Cal rerun rows.
• manifests/overlap_map.jsonl: Full31/Table1 overlap map over media_id.
• checksums/media_files.sha256: registry-derived media checksums.
• checksums/package_files.sha256: package metadata checksums.
• metadata/handoff_receipt.json: build receipt and human next steps.

Counts

• Unique included media files: 17642.
• Full31 included logical rows: 6200.
• Table1 total rerun rows: 14240.
• Table1 included media-backed rows: 14198.
• Table1 excluded denominator-caveat rows: 42.
• Full31/Table1 overlap groups: 80.

Boundaries

This package does not contain raw PAWEval provider responses, paid API logs, secrets, temporary runner directories, shard/chunk provenance, raw eval dumps, datasets, checkpoints, or generated job payload directories.

Paid PAWEval execution and Hugging Face upload are human-operated gates. The agent prepared this package and validation receipts only.