PhysEdit/pawbench-gemini-expansion-20260619

PAWBench Gemini Expansion 20260619

This dataset package contains the 5120-row PAWBench Gemini 3.5 Flash expansion input package for collaborator-side API evaluation.

Scope

• Cosmos3 Nano, Table 1 PAWBench main only: 1560 videos.
• Cosmos3 Super I2V, Table 1 PAWBench main only: 1560 videos.
• Latest clean LTX2.3, Table 1 PAWBench main: 1560 videos.
• Latest clean LTX2.3, K-scaling: 300 videos.
• Latest clean LTX2.3, explicit endpoint: 140 videos.

LTX causal/cue latest-clean surfaces are intentionally excluded for this package because no formal denominator is materialized in this sprint.

Layout

• videos/: materialized generated videos.
• source_frames/: first-frame inputs referenced by each row.
• manifests/all_rows_redacted.jsonl: row-level metadata with public Hugging Face paths and content checksums.
• instructions/gemini_api_input_final.jsonl: Gemini API input rows.
• instructions/friend_return_contract.md: expected raw/parsed/metrics return contract.
• checksums/videos.sha256 and checksums/source_frames.sha256: verification ledgers.

Evaluation

For every row in instructions/gemini_api_input_final.jsonl, call Gemini 3.5 Flash with the listed video_uri, source frame, action prompt, outcome menu, invalid criteria, and expected JSON schema. Store raw responses outside this repo first, then return parsed rows and metrics according to the contract.

Metadata correction 2026-06-22

The LTX2.3 K-scaling rows now carry outcome_menu values propagated from their referenced PAWEval outcome rubrics. The correction changes metadata only: row ids, media paths, videos, source frames, and the 5120-row denominator are unchanged.

K-scaling scoring note 2026-06-22

K-scaling is reported as a distribution-diagnostic surface. It still uses the same scene outcome menus as the PAWEval rubrics, and the collaborator should return a menu-constrained selected_outcome for each row unless the row is explicitly invalid.