LCZZZZ/proactive-eval-reviewed-20261019

Proactive eval review snapshot (2026-10-19)

🆕 NEW in this upload (2026-10-27): 2,000 extra UNREVIEWED candidates

	CT-G	CT-L	CT-M	PS-R
Newly added samples	500	500	500	500

• Everything else in this repo (the 5,742 qualified samples and the 700 pending PS-P) is unchanged from the 2026-10-19 snapshot.
• The new batch lives only in: data/extra_candidates_2000.jsonl (config extra_candidates), media/CT_ext/, media/PS_R_ext/, media_index/CT_ext.jsonl, media_index/PS_R_ext.jsonl.
• These samples have not passed human review and are not counted in the qualified tables below. Details: Extra candidates.

This snapshot contains 5,742 qualified samples from the CT, CM/MC/RUC, PS and JOY review sets, plus 700 newly selected PS-P candidates that are still unreviewed. The pending candidates are excluded from the qualified split. The video bundle contains 54 compressed volumes (about 143 GB total).

2026-10-27 update: added 2,000 extra unreviewed candidates (CT-G / CT-L / CT-M / PS-R, 500 each) in a separate split and media bundle; see Extra candidates. They are not part of the qualified counts below.

Counts

Dataset	Subclass	Qualified	Pending review
CT	G	794	0
CT	L	654	0
CT	M	531	0
CT	T	837	0
CM	G	143	0
CM	L	74	0
CM	M	122	0
CM	T	161	0
MC	G	153	0
MC	L	97	0
MC	M	133	0
MC	T	117	0
RUC	G	175	0
RUC	L	83	0
RUC	M	81	0
RUC	T	161	0
PS	A	24	0
PS	I	445	0
PS	P	119	700
PS	R	240	0
JOY	C	111	0
JOY	E	382	0
JOY	N	105	0
Total		5,742	700

Files

• data/reviewed_qualified.jsonl: one row per approved sample; sample contains the original annotation and review contains the reviewer-approved text, timing edits and audit metadata.
• data/ps_p_pending_700.jsonl: additional PS-P candidates with review_status=unreviewed. These have not passed human review.
• media/<group>/part-*.tar.zst: zstd-compressed tar volumes containing the original video files.
• media_index/<group>.jsonl: maps every sample_id to its archive, member path, and uncompressed video size. Use this index to locate a sample's video.
• data/extra_candidates_2000.jsonl: the 2,000 extra unreviewed candidates (same row schema as above, review_status=unreviewed, review=null).
• media/CT_ext/, media/PS_R_ext/ and media_index/CT_ext.jsonl, media_index/PS_R_ext.jsonl: videos and lookup for the extra candidates.
• counts.json: machine-readable subclass counts.

To extract one archive: tar --use-compress-program=zstd -xf media/CT/part-000.tar.zst. The paths inside the archive match the member field in media_index. Video files can be shared by multiple samples.

The review snapshot intentionally keeps original source fields for provenance, including source URIs and local paths. For this HF copy, use media_index to locate the bundled video. Source dataset terms may differ by source; check the license and source_dataset fields before redistribution.

Provenance

• CT: ct_criterion_review/reviews/Jiheng_Hou (G/L) and reviews/cy (M/T).
• CM/MC/RUC: cm_mc_ruc_review/reviews/cy.
• PS: ps_joy_review/reviews/Jiheng_Hou.
• JOY: ps_joy_review/reviews/cy.
• Pending PS-P: ps_joy_review/data/ps_p_expansion_700.jsonl, selected from the original PS40K pool.

Extra candidates (unreviewed, 2026-10-27) — NEW in this upload

2,000 additional candidates for the subclasses that needed more samples. None has passed human review; annotations are pipeline-generated (CT: Gemini-built Criterion_Timing; PS-R: Gemini-generated with automated verification, automated_status=passed). They are kept out of qualified.

Dataset	Subclass	Samples	Unique videos	Source datasets	Max share of one source	Where drawn from
CT	G	500	500	32	30 (6%)	Criterion_Timing pool, balanced split
CT	L	500	500	17	45 (9%)	Criterion_Timing pool: 406 balanced + 94 extra
CT	M	500	500	16	41 (8%)	Criterion_Timing pool: 389 balanced + 111 extra
PS	R	500	500	26	33 (6.6%)	ProactiveStreaming40K: 121 internal_test + 138 development + 241 train
Total		2,000	2,000

PS-R task mix: once_per_event_instance 240, every_occurrence_response 152, unique_target_counting 108 (same proportions as the qualified PS-R set).

Selection rules

• Every candidate is absent from the earlier review pools (including rejected items) and from the qualified and pending-PS-P sets above, by sample id, video name and video sha256.
• One sample per video, and no two samples in a class share a near-identical scene/task description (token-overlap filter).
• Samples are drawn round-robin across source datasets so no source dominates. CT-L and CT-M have only 17 and 16 usable sources, which is why their largest source reaches 8-9%.
• Prefers surplus eval data (internal_test / development / extra) over training data where available.

To locate a video, use media_index/CT_ext.jsonl (CT) or media_index/PS_R_ext.jsonl (PS-R); archives are media/CT_ext/part-*.tar.zst and media/PS_R_ext/part-*.tar.zst. PS40K sources carry license_status=review_required; check licenses before redistribution.