jimchen2/pendulum-causal-qa-3000

Pendulum Causal Video QA — 3K trajectories

A synthetic, simulator-grounded video question-answering dataset built from the nonlinear pendulum renderer and 160-frame setup in jimchen2/pendulum-cjepa-160f. The renderer/source generator is included; this repository extends it with causal QA annotations and OOD validation strata.

What is in this repository

• 42,000 QA records for 3,000 planned trajectories: 14 questions per video, with machine-readable answers, timestamps, evidence frames, and counterfactual intervention metadata.
• 33,600 QA rows for 2,400 training videos and 8,400 QA rows for 600 validation videos.
• data/train.jsonl, data/validation.jsonl: one QA object per line, partitioned by whole video.
• data/questions.csv: flat copy of all QA rows.
• data/index.csv: one row per planned video, with physical parameters, split, and OOD group.
• samples/12_video_demo/: a small, fully rendered 12-video example (160 frames at 12 fps), frame-aligned CSV labels, and 168 questions. It includes one validation example from each OOD group.
• generate_causal_qa_dataset.py, pendulum_cjepa.py, verify_dataset.py, and requirements.txt.

Important media note: the 3,000 full MP4s are not hosted in this repository. Only the 12 rendered demo clips are included. The complete QA annotation pack and parameter index are present; the included generator can render all 3,000 clips and matching CSVs/QA locally.

Load the questions

from datasets import load_dataset

qa = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/validation.jsonl",
    },
)
print(qa["train"][0])

The JSONL fields include question_id, video_id, video_path, split, ood_group, question_type, question_family, question, answer, answer_type, answer_value, time_s, evidence_frames, query, and intervention. video_path points to the full generated dataset layout; for this hosted repository, use the clips under samples/12_video_demo/ as examples.

Split protocol: OOD validation, not a seeded random holdout

The 2,400 training trajectories are sampled from the source in-distribution ranges: initial angle magnitude 15–40°, initial angular velocity in [−0.20, 0.20] rad/s, damping 0.15–0.50 s⁻¹, light mean 70–110°, light amplitude 0–15°, and light frequency 0.04–0.12 Hz. Training uses gravity 9.81 m/s² and length 1.0 m.

The 600 validation trajectories are held out by OOD regime, with 100 videos per group:

1. val_ood_amplitude: initial angle magnitude 45–62°.
2. val_ood_near_conservative: damping 0–0.12 s⁻¹, including undamped examples.
3. val_ood_initial_velocity: angular velocity magnitude 0.35–0.60 rad/s.
4. val_ood_gravity_length: unseen lengths (0.65–0.80 m or 1.25–1.55 m) or gravity values (7.5–8.5 or 11.5–12.5 m/s²).
5. val_ood_illumination: held-out light means and frequencies.
6. val_ood_compound: multiple held-out physical factors combined.

The validation schedules are deterministic, low-discrepancy parameter designs independent of the training --seed. The split is by complete video, not by adjacent frames. See data/index.csv and data/metadata.json for per-video labels.

Question types and causal scope

Each video has 4 descriptive, 2 predictive, 2 explanatory, and 6 counterfactual questions. Topics include bob side/direction, speed comparisons, center crossings, future state, energy and damping, light's lack of causal effect on bob dynamics, and re-simulated interventions on damping, initial conditions, length, and gravity. Answers are calculated from RK4 simulator trajectories rather than generated by a language model.

This is CLEVRER-inspired, but it is not a multi-object collision dataset. The underlying simulator contains one pendulum bob, a rod, a changing light, and a shadow; it does not model collisions or interactions between multiple objects. The causal claims are specific to the simulator equations.

Render the full 3,000-video dataset

python -m pip install -r requirements.txt
python generate_causal_qa_dataset.py \
  --output ./pendulum_causal_qa_3000_full \
  --num-videos 3000 --train-fraction 0.8 --seed 42

This writes videos/train/, videos/val/, frame-aligned timeseries/, index.csv, and the full questions/answers. The settings above reproduce the IDs and QA records hosted here. If interrupted, rerun the same command with --resume.

License and provenance

The pendulum renderer/simulator is adapted from jimchen2/pendulum-cjepa-160f; the upstream MIT license is preserved in LICENSE.