hunarbatra/4DReasoner_v5_train_100k

4DReasoner v5 — Training Mix

A curated multi-source training corpus for 4D (spatial + temporal) video-language reasoning. Two sizes are released:

• 4DReasoner_v5_train_20k — 20,000 rows, used for fast SFT iteration / ablation runs.
• 4DReasoner_v5_train_100k — 99,500 rows (≈100K), used for final-method runs.

The mix is bench-aligned to seven evaluation benchmarks: VSI-Bench, VSTI-Bench, STI-Bench, DSR-Bench, DSI-Bench, VLM4D real, and VLM4D synthetic.

Mix composition

Five source families, weighted by overlap with the seven evaluation benchmarks. The 20K preset preserves the same source ratios as the 100K, just at 1/5 scale.

Source	20K	100K	What it covers
VLM-3R VSTI-Bench train (Journey9ni/VLM-3R-DATA)	7,000	35,000	Camera motion direction, object↔camera distance, object↔object position. Surface-form match for VSTI-Bench eval.
VLM-3R VSI-Bench train (Journey9ni/VLM-3R-DATA)	6,000	30,000	Distances, object size, route planning, room-scale spatial relations. Surface-form match for VSI-Bench eval.
SpaceR-151k spatial-video (RUBBISHLIKE/SpaceR-151k)	3,000	20,000	General spatial reasoning over real video (LLaVA-Video / STAR / CLEVRER subsets).
RoboFAC (MINT-SJTU/RoboFAC-dataset)	2,000	10,000	Robotic-arm action understanding from stable-camera videos. Matches the 4D-RGPT (CVPR'26 Highlight) training mix.
SAT-Spatial-VQA (hunarbatra/SAT-Spatial-VQA-5k)	2,000	4,500	Image-based dynamic-scene spatial reasoning. Style transferable to DSI character_motion.
Total	20,000	99,500

Eval-set deduplication

Every row is checked against a 312-scene blocklist drawn from the union of:

• VSTI-Bench test.json ScanNet eval scenes
• VSI-Bench test.jsonl ScanNet eval scenes

Train rows whose source video is in the blocklist are excluded, so no eval-set contamination flows into training.

Schema (30 columns, v3 parity)

The dataset matches the hunarbatra/4DReasoner_v3_full parquet schema. Headline columns:

• question (str), options (list[str]), correct_letter (str|None)
• category (str), source_dataset (str), scene_name (str|None)
• video_path (str|None) — relative path to source video (videos must be downloaded separately from each source's HF repo)
• sharegpt_no_cot (list[dict]) — three-turn conversation: system, human, gpt-letter
• sharegpt_cot_qual / sharegpt_cot_quant — None for now; will be filled by the 4DReasoner v5 fact-pack + CoT pipeline
• 4D_facts, extracted_4d_data, cited_fact_id, candidate_facts, context_fact_ids — None for now; reserved for fact-pack injection
• reasoning_qual, reasoning_quant, cot_gen_prompt_qual, cot_gen_prompt_quant — None for now; reserved for CoT generation
• frame_paths, frame_indices, timestamps, frame_sample_fps, frame_type, caption — None for now; reserved for L4P-extracted geometry
• prompt_version = "v5_curated", context_mode = "no_context"

Intended use

Stage-1 SFT of Qwen3-VL-8B (and similar VLMs) on bench-aligned spatio-temporal QAs. The reserved columns are filled by later v5 stages:

1. Stage 1 SFT — train directly on sharegpt_no_cot (this dataset, as-is).
2. Stage 2 fact-pack extraction — run L4P (depth + flow + dynamic mask + 3D tracks + camera pose) on each video, render structured facts, fill the 4D_facts / extracted_4d_data / cited_fact_id columns.
3. Stage 3 CoT generation — prompt an LLM teacher with (question, answer, fact_pack), get back reasoning_qual / reasoning_quant.
4. Stage 4 latent alignment — distill L4P features into the VLM via a 3-layer MLP projector head (4D-RGPT pattern).
5. Stage 5 OPCD — on-policy context distillation: teacher with fact pack → student without.

Allocation rationale

• VLM-3R is dominant (~65%) because its source domain (ScanNet / ScanNet++) matches the VSI / VSTI eval benchmarks and the QA surface form is identical.
• RoboFAC + SpaceR (~25-30%) add diversity: stable-camera object-motion + general spatial reasoning, both of which the eval benchmarks indirectly test.
• SAT (~5-10%) adds image-based spatial reasoning — small but covers a reasoning style the video-only sources don't.
• 20K is for fast iteration: pre-method-finding SFT runs to identify what works (latent alignment vs CoT-SFT vs OPCD vs combinations).
• 100K is for final runs once the method is locked.

Sources not included (and why)

• Wolf NuScenes (autonomous driving captions, used by 4D-RGPT) — not available on HF; would require GitHub project-page download. Marked TODO.
• ST-VLM / STKit (kinematic instruction tuning, ikodoh.github.io/ST-VLM) — not on HF. Marked TODO.
• EmbodiedScan (cjfcsjt/embodiedscan) — 3D occupancy + visual grounding, not video QAs. Reserved for the latent-alignment phase as a 3D supervision signal, not for SFT MCQs.

Citation

If you use this dataset, please cite the originating sources:

@inproceedings{vlm3r2026,
  title={VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction},
  author={...},
  booktitle={CVPR},
  year={2026}
}
@misc{spacer2025,
  title={SpaceR: Reinforcing MLLMs in Video Spatial Reasoning},
  year={2025},
  archivePrefix={arXiv},
  eprint={2504.01805}
}
@misc{robofac2025,
  title={RoboFAC},
  year={2025}
}
@misc{sat,
  title={SAT: Spatial Aptitude Training for Multimodal Language Models}
}