Datasets:

Randallhy
/

RefineCut-Bench

license: cc-by-nc-4.0
language:
  - en
task_categories:
  - text-generation
tags:
  - video-editing
  - planning
  - verifier
  - benchmark
pretty_name: RefineCut-Bench
size_categories:
  - 1K<n<10K

RefineCut-Bench

A planning-level benchmark for executable video-editing planning (EMNLP 2026, Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing). A task gives a planner a brief, a real clip pool with schema-constrained captions and metadata, optional music metadata with beat tracks, the current timeline state, and an explicit constraint ledger; the planner emits a RefinePatch (RFC 6902-style JSON Patch over a typed timeline) and a deterministic verifier applies it and recomputes every ledger entry. Paper: https://arxiv.org/abs/2608.25622 (EMNLP 2026 Main). Evaluation code and prompt: https://github.com/Lancelot-wy/RefineCut.

	Count
Canonical tasks (raw task records)	3,578 (3,960)
Task families / subtypes	3 (A composition 1,460 · B local edit 1,219 · C generative 899) / 15
Constraint types	14
Captioned clips	7,971
Music tracks (features only)	499
Train / dev / test (task-id disjoint)	2,773 / 596 / 591
Evaluation sets	Common-100, dev100, canonical-clean (92)
Teacher trajectories, replayed	GPT-5.4 2,000 · Qwen3-Max 1,946 · DeepSeek-V4-Pro 1,959

Layout

tasks/        all_tasks.jsonl, train.jsonl, dev.jsonl, test.jsonl, split_manifest.json, split_audit.json
eval/         common100_items.jsonl   evaluation-ready Common-100 items (ledger, clip-pool metadata, failed
                                      initial state, violated constraints) consumed by the evaluation harness
              dev100_items.jsonl      checkpoint-selection set
              test_items.jsonl        all 591 test items in the same format
              canonical_clean_ids.json   the 92 Common-100 tasks whose canonical id never appears in training
clips/        captions.jsonl (7,971 clips: subject / action / scene / camera / scene_category / motion_intensity /
              caption_short / duration / source), clip_alias_maps/, clips_meta/<source>/ (public source identifiers)
music/        music_features.json (BPM, beat times, energy, duration for 499 tracks; no audio)
trajectories/ normalized/<teacher>.jsonl     canonicalized multi-teacher trajectories (up to 3 steps x 4 branches)
              replayed/<teacher>_replay.jsonl  the same trajectories with per-branch verifier replay scores
schemas/      constraint_ledger, editplan, refinepatch, timeline_ir, verifier_output JSON schemas
docs/         metric definitions and the VES formula

Task record

{"task_id": "...", "task_type": "A", "task_subtype": "themed_montage", "brief": "...",
 "target_duration": 24.0,
 "constraint_ledger": [{"item_id": "...", "type": "must_keep_clip", "spec": {...}, "satisfied": false, "evidence": null}, ...],
 "clip_pool": ["clip_0001", ...], "structure": {...}, "canonical_id": "..."}

Constraint types: must_keep_clip, must_exclude_clip, target_duration, duration_tolerance, pacing, transition_style, music_sync_bpm, music_sync_beat, must_open_with, must_close_with, no_repeat_within_seconds, max_repeats_per_clip, tag_inclusion, tag_exclusion. An entry is a hard constraint when its spec admits a deterministic pass/fail test; softer entries earn graded credit through the fractional constraint-satisfaction rate.

Evaluation protocol

Closed loop with at most T=3 repair steps, greedy decoding, one fixed PatchPlanner prompt, and the frozen Common-100 list. VES = 0.30 FinalCSR + 0.15 HardPass + 0.15 PASR + 0.15 ReqClipRecall + 0.10 DurationPass + 0.10 TimelineValidity + 0.05 NoRegression (docs/METRIC_DEFINITIONS.md). VES is a protocol-specific executable-planning score; rendered video quality is evaluated separately.

Sources and license

Clips: Pexels (3,219), Panda-70M sample (3,990), Pixabay (255), OpenVid (100), and 5-30 s segments of non-game long videos (407); music features from FMA tracks. Raw video and audio are not redistributed; clips/clips_meta/ gives the public source identifiers. Benchmark metadata, captions, ledgers, and trajectories: CC BY-NC 4.0; sources retain their original licenses.

Citation

@inproceedings{refinecut2026,
  title     = {Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing},
  author    = {Wang, Haoyu and Feng, Cheng and Bian, Liuyang and Huang, Ruiyang and Wei, Lei and Wen, Yafei and Chen, Xiaoxin and Tang, Xiaoying},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026},
  note      = {to appear}
}