videorl-video-val

5,645-row full-set video evaluation suite spanning three public benchmarks: Video-MME, PerceptionComp, and Video-Holmes. Used as the offline evaluation set for the multimodal-rlvr video RLVR project — reported as the "full-set offline val" in all 8B video RL run post-mortems.

VIDEOS ARE NOT INCLUDED. Parquets store an external file:// path reference to each video. You must download the original public datasets and update paths to match your local layout. See the section below.

---

Benchmark breakdown

Config / parquet	data_source value	Rows	What it tests
videomme-test.parquet	videomme_v1_test	2,700	Broad video understanding: perception, cognition, temporal reasoning, domain knowledge; multiple durations (short/medium/long)
perceptioncomp-test.parquet	perceptioncomp_test	1,108	Compositional perception and fine-grained scene understanding in video
videoholmes-test.parquet	video_holmes_test	1,837	Causal and counterfactual video reasoning; "whodunit"-style inference
Total		5,645

All rows are multiple-choice with single-letter ground truths.

---

Frame and pixel configuration

Val parquets are curated at higher resolution than the 24f100k training set, to better reflect benchmark conditions:

Parameter	Value
fps	1
max_frames	64
min_pixels	3,136 (56×56)
max_pixels	501,760

---

Schema

Identical schema to the training set. Each row contains:

Column	Type	Description
prompt	list[{role: str, content: str}]	Chat-formatted prompt. system message contains the reasoning format instruction. user message contains <video> placeholder + MCQ question.
images	list[{bytes, path}]	Always empty for video rows.
videos	list[{type, video, fps, max_frames, min_pixels, max_pixels}]	One entry. video is an external file:// path — not embedded bytes.
data_source	str	Benchmark identifier: videomme_v1_test, perceptioncomp_test, or video_holmes_test.
ability	str	Always video_understanding.
reward_model	{ground_truth: str, style: str}	ground_truth is a single letter (A–D). style is rule.
extra_info	struct	reward_type (multiple_choice), answer, plus benchmark-specific metadata fields (e.g. task, category, question_type, subject).

Extra_info metadata fields (val-only)

The val parquets carry richer per-row metadata inherited from the source benchmarks. Not all fields are populated for all sources — unset fields are null.

Field	Populated by
task	VideoMME (e.g. Counting Problem, Action Recognition)
category	VideoMME (e.g. Humanity & History, Science)
question_type	VideoMME (short, long)
subject	VideoMME, PerceptionComp
source_id	PerceptionComp, Video-Holmes

Prompt system message (baked in)

Every row embeds the same reasoning system prompt as the training set:

You are a careful reasoning assistant. ALWAYS respond in this EXACT format:

<think>step-by-step reasoning</think>
<answer>\boxed{final_answer}</answer>

Example row (abbreviated — VideoMME)

{
  "prompt": [
    {"role": "system", "content": "You are a careful reasoning assistant. ALWAYS respond in this EXACT format:\n\n<think>step-by-step reasoning</think>\n<answer>\\boxed{final_answer}</answer>\n..."},
    {"role": "user",  "content": "<video>\nHow many times does the presenter point at the whiteboard?\nA. Once\nB. Twice\nC. Three times\nD. Four times\nPlease reason step by step, then provide your final answer."}
  ],
  "images": [],
  "videos": [
    {
      "type": "video",
      "video": "file:///YOUR_CACHE_ROOT/videomme/data/fFjv93ACGo8.mp4",
      "fps": 1,
      "max_frames": 64,
      "min_pixels": 3136,
      "max_pixels": 501760
    }
  ],
  "data_source": "videomme_v1_test",
  "ability": "video_understanding",
  "reward_model": {"ground_truth": "C", "style": "rule"},
  "extra_info": {
    "reward_type": "multiple_choice",
    "answer": "C",
    "task": "Counting Problem",
    "category": "Humanity & History",
    "question_type": "short"
  }
}

---

Videos are NOT included — how to obtain them

The video field is an absolute local path. You must download the source videos and remap the path prefix to your local storage.

Per-benchmark video download

data_source	Embedded path pattern	Public source	How to download
videomme_v1_test	file:///YOUR_CACHE/videomme/data/<video_id>.mp4	Video-MME	lmms-lab/Video-MME — download the data/ split; videos are MP4s named by YouTube ID
perceptioncomp_test	file:///YOUR_CACHE/datasets--hrinnnn--PerceptionComp/snapshots/<hash>/...	PerceptionComp	hrinnnn/PerceptionComp via huggingface_hub or datasets library
video_holmes_test	file:///YOUR_CACHE/datasets--TencentARC--Video-Holmes/snapshots/<hash>/...	Video-Holmes	TencentARC/Video-Holmes via huggingface_hub or datasets library

After downloading, update the video paths in the parquets to match your layout before running evaluation:

import pyarrow as pa, pyarrow.parquet as pq

OLD_PREFIX = "file:///scratch/users/astar/ares/suns1/trung/cache"
NEW_PREFIX = "file:///YOUR/LOCAL/PATH"

for fname in ["videomme-test.parquet", "perceptioncomp-test.parquet", "videoholmes-test.parquet"]:
    t = pq.read_table(fname)
    rows = t.to_pylist()
    for r in rows:
        for v in r.get("videos") or []:
            v["video"] = v["video"].replace(OLD_PREFIX, NEW_PREFIX)
    pq.write_table(pa.Table.from_pylist(rows, schema=t.schema), fname)

---

Reward and answer format

• Reward type: multiple_choice for all three benchmarks
• Ground truth: single uppercase letter (A, B, C, or D)
• Expected model output format:

  <think>...chain-of-thought reasoning...</think>
  <answer>\boxed{C}</answer>
  

• Accuracy rule: extract letter from \boxed{...} inside <answer> block; fall back to lmms-eval's extract_mcq_answer priority-ranked extractor.
• Format reward: 1.0 iff <think>...</think><answer>...</answer> tags present with ≥ 100 non-whitespace chars in <think>; 0.0 otherwise.
• Composite score used in reported results: 0.8 × accuracy + 0.2 × format

---

Reported results (multimodal-rlvr 8B video RL campaign)

Best checkpoints across all 8B video RL runs evaluated on this full set:

Run	Best step	Val accuracy
Cold-start base RL (4-node)	global_step_80	~0.4918
SFT-770 warm-start RL (4-node)	global_step_180	~0.4845
Base + ppexplore τ0.95	global_step_140	~0.489
SFT-770-RL60 + ppexplore τ0.95	global_step_20	~0.4881
SFT-770-RL60 + ppexplore τ0.8	global_step_20	~0.4864
SFT-770-RL60 + ppexplore τ0.8, n16k8	global_step_80	~0.4913

All runs plateau at ~0.485–0.492 (vs 0.4444 for untuned base), a consistent +4–5 pt gain from RL, with no lever breaking the ceiling across 7 experiments.

---

Curation notes

• VideoMME: full test split (2,700 rows = 900 questions × 3 video durations: short/medium/long).
• PerceptionComp: test split filtered to cached video files (1,108 rows).
• Video-Holmes: full test split (1,837 rows).
• All curated with scripts/data/curate_v1.py val-canonical-video from the multimodal-rlvr project.
• Videos are decoded at 1 fps, up to 64 frames, via qwen_vl_utils during inference.

---

Licenses

Source benchmarks retain their original licenses:

Benchmark	License / Terms
Video-MME	CC BY 4.0
PerceptionComp	Apache 2.0
Video-Holmes	Apache 2.0

This repository (parquet curation structure) is released under Apache 2.0.

---

Citation

If you use this evaluation set, please cite the original benchmarks:

@inproceedings{videomme,
  title={Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis},
  author={Fu, Chaoyou and others},
  booktitle={CVPR},
  year={2024}
}

@article{perceptioncomp,
  title={PerceptionComp: A Comprehensive Benchmark for Compositional Video Perception},
  author={Huang, Rui and others},
  journal={arXiv preprint},
  year={2024}
}

@article{videoholmes,
  title={Video-Holmes: Can MLLMs Think Like Detectives?},
  author={TencentARC},
  journal={arXiv preprint},
  year={2024}
}