TrajQA-B — metric motion questions on sparse 4D trajectories

Data release of Sparse Trajectory Tokens for Metric Motion Queries in Multimodal Language Models (ICLR 2027 submission). The code (data pipeline, tokenizer, training, evaluation, figure scripts) is in the accompanying repository; the model weights are in the companion model repo SparseTrajTokens.

Everything here is derived from Stereo4D (YouTube VR180 videos and their released 3D tracks). Video frames and raw videos are not redistributed; the code repository extracts them from the official release deterministically (frames re-extracted this way are byte-identical to the ones we trained on).

Contents

Path	Size	Content
trajqa_b/final/	68 MB	TrajQA-B: train.jsonl 56,120 / val.jsonl 1,318 / test.jsonl 4,492 / heldout_test.jsonl 3,430 questions; _audit.json (text-shortcut audit), _heldout_test_report.json
trajqa_b/candidates_m6/, trajqa_b/candidates_m1/	360 MB	Generator candidates (m6: 333,661) and their Gate-0 token-side recomputation (m6 → train / held-out, m1 → val / test)
trajqa_b/build_m1/	23 MB	Intermediate default build (val/test identical to final/; its train split is the family-proportion reference)
captions/t8_train.jsonl	27 MB	65,556 rule-generated motion captions used for Stage 1 (42,030 scenes)
tokens_v4td.tar (+ tokens_v4td_meta.json)	1.3 GB	Trajectory tokens of the 42,151 scenes (tokens_v4td/scene_<clip>.pkl, pickled SceneTokens; needs src/tokenization/schema.py on sys.path)
scenes/scenes-NN.tar (64 shards) + scenes/INDEX.tsv	108 GB	Processed scenes scene_<clip>/: final_repr.pkl (sparse metric keypoint tracks per object), dense_tracks_<iid>.npz (released tracks attributed to each instance), stage0_pre.json (VLM object inventory), rigid_route.json, gt_dense_stats.json, scene_manifest.json, stereo4d_meta.json, gt_stereo4d.npz (camera poses / timestamps only), … INDEX.tsv maps each scene to its shard
figure_scenes.tar	2.7 GB	SAM 3 re-run (all frames) for the 22 clips shown in the paper figures and the demo
curation/	21 MB	Outcome of every Stereo4D clip (108,015 → 42,152): delivered list, rejection reason per clip, rejected-clip evidence (skipped_static_records.tar.gz)
results/	30 MB	Final per-question evaluation outputs, checkpoint-selection evaluations on the validation split, training logs (meta.json, train_log.jsonl, val_curve.jsonl, PREREG.md), tracker bake-off scoreboard
baselines/	47 MB	Per-question predictions of the 15 baseline models (preds/), the evaluation questions in baseline format (qa/), result tables and figures
paper_assets/	250 MB	Rendered figure panels and demo videos
MANIFEST.tsv		SHA-256 of every tar

Question format

One JSON object per line: id, video_id (= scene), event_subtype (family: cam_disp_grid, disp_grid, speed_grid, interval_disp, interval_speed, motion_source, motion_duration_binary), question, options, answer, evidence (structured target), frame_range, gt_payload (reference value, grid window and cell, Gate-0 recomputation, …).

Splits are formed on a hash of the source YouTube video id: test u < 0.20, val [0.20, 0.25), held-out [0.25, 0.30), train otherwise — the four files share no source video.

Unpacking

mkdir -p data && cd data
tar xf tokens_v4td.tar && tar xf figure_scenes.tar
mkdir scenes && for t in scenes/scenes-*.tar; do tar xf "$t" -C scenes; done      # 108 GB
sha256sum -c <(awk -F'\t' '{print $3"  "$1}' MANIFEST.tsv)                        # optional

The resulting data/ directory matches the layout expected by the code repository (data/README.md there).

Notes

• The Stage-1 caption corpus was not built with the QA splits: its 64,523 training captions cover 2,891 of the 3,103 QA evaluation scenes (question-free descriptions of displacement and speed from the same tracks).
• Reference answers inherit the errors of the Stereo4D reconstruction (tracks, camera poses, masks); see the paper's limitations section.
• License: CC BY 4.0, the same terms as the Stereo4D annotations this release is derived from (https://creativecommons.org/licenses/by/4.0/). Please cite Stereo4D (Jin et al., 2025) together with our paper. The underlying videos belong to their YouTube creators and are not part of this release.