ngqtrung/hopchain-v40-1080p

HopChain v4.0 + Vript — Harder Multi-Hop Video Reasoning with Verifiable Answers

28,865 multi-hop video questions over 7,396 videos, each with an integer ground-truth answer checkable by rule (no LLM judge at train time). Built for RLVR training.

This dataset combines two sources:

• HopChain v4.0 (6,692 QAs / 1,748 videos): 1080p caption-grounded, conditional-predicate multi-hop questions over diverse YouTube videos
• Vript (22,173 QAs / 5,648 videos): dense scene-annotated videos (>3min), re-captioned with 397B and processed through the same HopChain pipeline

Both sources use the same question format, system prompt, and reward contract.

Splits

Split	Rows	HopChain	Vript	Videos	Hops (3/4/5/6/7)	Solve Rate
Train	25,975	6,004	19,971	6,657	523 / 4,957 / 9,438 / 8,013 / 3,044	0.0081
Val	2,890	688	2,202	739	46 / 557 / 1,071 / 879 / 337	0.0088
Total	28,865	6,692	22,173	7,396	569 / 5,514 / 10,509 / 8,892 / 3,381	0.0082

Video-disjoint split (seed=42): no video appears in both train and val.

What is HopChain v4.0

HopChain v4.0 is the difficulty successor to v3.4. Each question chains 3–6 conditional-predicate hops (mode 4–5). A hop observes one robust fact about a scene and a binary rule turns it into a number — if <fact> then A else B — aggregated with + * into a single verifiable integer.

11 reasoning families

• Original (6): existence, color, action, cooccur, setting, arithmetic
• New cross-scene (5): order, first_last, attr_match, recurrence, state_change

Difficulty gate

Too-easy probe = training model (Qwen3-VL-8B-Instruct, 140 frames), dropped if solve_rate ≥ 6/8.

What is Vript

Vript provides ~11K high-quality YouTube videos with dense scene-level annotations (shot type, camera movement, content, voiceover). We use 8,046 videos >3 minutes (median 7m12s, mean 9m), re-captioned with 397B and processed through the same HopChain pipeline.

Vript's longer videos (3–7 min vs 20–60s for HopChain) provide richer temporal reasoning opportunities.

Schema

Each row is a complete verl-compatible prompt:

{
  "prompt": [
    {"role": "system", "content": "You are a careful reasoning assistant..."},
    {"role": "user", "content": "<video>\n<question>"}
  ],
  "images": [],
  "videos": [{"type": "video", "video": "file:///path/to/video.mp4", "fps": 1,
              "max_frames": 140, "min_pixels": 3136, "max_pixels": 50176}],
  "data_source": "video_multihop_prototype",
  "ability": "video_multihop_reasoning",
  "reward_model": {"ground_truth": "42", "style": "rule"},
  "extra_info": {
    "reward_type": "numeric",
    "answer": "42",
    "answer_type": "numeric",
    "num_hops": 5,
    "hop_types": ["order", "cooccur", "action", "color", "arithmetic"],
    "question_id": "VIDEOID_hq000",
    "video_id": "VIDEOID",
    "solve_rate": 0.0,
    "difficulty_band": "hard"
  }
}

Key fields

• reward_model.ground_truth: integer answer (exact-match reward)
• extra_info.num_hops: number of conditional-predicate observation hops
• extra_info.hop_types: list of reasoning types per hop
• extra_info.solve_rate: fraction of 8 runs where training model (Qwen3-VL-8B-Instruct, 140f) solved it — use for curriculum sampling
• extra_info.difficulty_band: "hard" (99.5%), "medium", "easy"

Video paths

• HopChain videos: file:///scratch/users/astar/ares/suns1/trung/videoreason/proxy_140f50k/{VIDEOID}.mp4 (140-frame proxies, 50176 max pixels)
• Vript videos: file:///scratch/users/astar/ares/suns1/tmp/vript/videos/{VIDEOID}.mp4 (original resolution)

Both use fps=1, max_frames=140, min_pixels=3136, max_pixels=50176.

Funnel

HopChain v4.0

Stage	Count
Input videos	1,818
Videos processed	1,787 (98.3%)
Total candidates	14,021
Judge survivors	7,539 (53.8%)
Final keepers	6,692 (3.74/vid)

Vript

Stage	Count
Input videos	8,046
Videos processed	~5,800
Final keepers	22,173 (3.82/vid)

Pipeline

1. Caption: 397B-FP8 generates dense captions from video frames (HopChain: 1080p oracle; Vript: re-captioned from scene annotations)
2. Generate: 397B-FP8 produces 5 candidate questions per video
3. Judge: 27B scores candidates, selects best
4. Too-easy filter: 8B-VL probes each question (8 runs), drops if solve_rate ≥ 6/8
5. Blind-solver filter: removes language-only shortcuts (rate 0.071)

Curriculum sampling

The solve_rate field enables curriculum training:

• Warm-up: rows where solve_rate > 0 (easier, 3–4 hop, simpler families)
• Anneal: into the 99.5% hard tail (solve_rate = 0, 5–7 hop)

Notes

1. Mixed caption oracle: HopChain videos: 144/1,818 captioned by 122B, rest by 397B-FP8. Vript: all re-captioned by 397B.
2. Generation never sees a frame: caption is the only ground truth. Blind-solver rate (0.071) = shortcut ceiling.
3. Video-disjoint split: ensures no video leakage between train and val.

Reproduction

Built from the VideoReasoningBench HopChain v4.0 pipeline.

Citation

@misc{hopchain2026,
  title={HopChain: Harder Multi-Hop Video Reasoning with Verifiable Answers},
  author={Trung, Ng Quang},
  year={2026},
  howpublished={\\url{https://huggingface.co/datasets/ngqtrung/hopchain-v40-1080p}}
}