jimchen2/pendulum-cjepa-3k-new

Pendulum C-JEPA 3K + CLEVRER-style Questions

3,000 silent H.264 MP4s of a nonlinear physical pendulum (160 frames @ 12 fps = 13.33 s, 196×196 px), with frame-aligned CSV labels and ~24,000 CLEVRER-style questions (descriptive / explanatory / predictive / counterfactual) whose answers are all derived from the ground-truth simulation.

This extends jimchen2/pendulum-cjepa-160f (300 videos) two ways:

1. 10× more videos, same renderer, same physics conventions, same CSV schema.
2. The validation split is out-of-distribution physics, not a held-out seed. Training videos all live in the original parameter box; every validation video breaks at least one physical assumption, so a model that merely memorizes the training distribution generically fails there. This matches the intent of a causal world-model dataset.

split	videos	physics
train	2,400	in-distribution (identical recipe to the original dataset)
val	600 (6 regimes × 100)	OOD — see regimes below

OOD validation regimes

regime	what changes vs train	why it is a causal test
val_high_damp	γ ∈ U(0.55, 0.90) s⁻¹ (train: 0.15–0.50)	envelope time-constant τ=2/γ as low as 2.2 s, shorter than anything seen
val_low_damp	γ ∈ U(0, 0.08) s⁻¹	near-conservative dynamics; amplitude/energy almost constant
val_large_angle	θ₀ = ±(45–80)° (train: ±15–40°)	strong nonlinearity: period elongated up to ~30%, visibly non-sinusoidal
val_gravity	g ∈ {1.62 Moon, 3.71 Mars, 24.79 Jupiter} m/s²	period shifts by up to 2.5× both directions
val_quad_damping	drag = −β·ω·|ω|, β ∈ U(0.10, 0.35); γ = 0	new force law: amplitude envelope decays ~linearly, not exponentially
val_driven	periodic torque A·cos(2πf t+φ), A ∈ U(0.35, 1.0) rad/s², f ∈ U(0.35, 0.65) Hz	new causal mechanism: beats / resonant growth around the ~0.5 Hz natural frequency; energy no longer decays monotonically

Config rejection-sampling guarantees every simulated frame stays inside the scene and |θ| ≤ 85° (the driven regime is the main source of rejections, near resonance).

Question set (CLEVRER taxonomy)

~8 questions per video, ~24,000 total. Format mirrors CLEVRER:

{
  "info": {...},
  "scene_annotation": [
    {"video_id": "video_00000",
     "video_filename": "videos/train/video_00000.mp4",
     "split": "train", "regime": "train",
     "questions": [
       {"question_id": 0, "question_type": "descriptive",
        "question": "...", "answer": "left"},
       {"question_id": 1, "question_type": "counterfactual",
        "question": "...", "choices": ["...", "..."], "answer": 0}
     ]}
  ]
}

• descriptive — observable facts: release side, bottom-crossing count, end-of-clip motion, damping envelope phrase, light drift, energy shape, fastest point, shadow direction. Mostly free-form string answers; three templates are convenience multiple-choice.
• explanatory (4-way MC) — what is responsible for what you see: amplitude decay (friction), stopping at the top (velocity through zero), fastest at the bottom (energy exchange), shadow sweeping (bob blocks light), never returning to release height (dissipated energy), beats in driven clips (rhythmic external push), large-angle period elongation (sin θ vs θ).
• predictive (4-way MC) — next peak higher/lower/same, motion right after the clip ends, bottom crossing within the next second, long-term fate (settle / keep swinging / driven pattern), future shadow motion. All computed from a hidden 28-second ground-truth continuation of the simulation (videos show only the first 13.33 s), so these have real answers, unlike CLEVRER's heuristic ones.
• counterfactual (4-way MC) — interventions: remove damping (returns to release height), add damping to an undamped clip, double gravity (faster), double rod length (slower), double bob mass (no change — mass independence), move/brighten the light (no effect on motion — the illumination is an independent nuisance factor by construction), steeper release (period slightly longer — amplitude coupling), double damping (dies out ~twice as fast), turn the drive off (settles), vacuum removes quadratic drag, Earth ↔ other-planet gravity.

Each question carries a concept tag (e.g. damping envelope, interventional invariance, g dependence of period) so you can score accuracy per causal concept and per OOD regime.

Layout

.
├── videos/train/video_00000.mp4 ...        2400 clips
├── videos/val/video_10000.mp4 ...           600 clips (val ids offset by 10000)
├── timeseries/video_XXXXX.csv              3000 frame-aligned label files
├── questions/questions_train.json          ~19k questions
├── questions/questions_val.json            ~4.8k questions
├── index.csv                               per-video params, split, regime
├── metadata.json
├── sample_sheet.png
└── scripts/
    ├── pendulum_gen.py                     physics + renderer + regime samplers
    ├── questions.py                        CLEVRER-style question generator
    ├── run_generate.py                     parallel generator (resume-safe)
    ├── verify_dataset.py                   integrity checks
    ├── audit_answers.py                    re-derive & audit question answers
    ├── make_preview.py                     build preview.html
    ├── prepare_cjepa_root.sh               recreate cjepa_data_root/ for extract_four.py
    └── upload_to_hf.py                     push this folder to the Hub

Frame n ↔ CSV row frame=n ↔ time_s = n/12. Splits are whole-video.

Physics

theta'' = -(g/L)·sin(theta) - gamma·theta' - beta·theta'·|theta'| + A·cos(2π f_d t + φ)

RK4 at fps × 8 = 96 Hz. Train: g=9.81, L=1.0, γ∈U(0.15,0.50), β=0, A=0, θ₀=±U(15,40)°, ω₀=U(−0.2,0.2), light mean U(70,110)°, light amp U(0,15)°, light freq U(0.04,0.12) Hz, random phase — identical to the original dataset. Illumination is independent of the dynamics (interventional-invariance property used by the counterfactual light questions).

energy_j_per_kg, theta_rad, omega_rad_s, alpha_rad_s2, bob/light/shadow geometry are all in the CSVs, so numeric extrapolation error at frames 128→159 (+2.67 s ≈ 1.3 oscillations) can also be scored exactly, as in the original.

Reproducing / extending

python scripts/run_generate.py --root dataset --n-train 2400 --val-per-regime 100 --seed 20261014
python scripts/verify_dataset.py dataset

Generation is seed-deterministic (per-video SeedSequence([seed, index])), runs at ~1 vid/s/core, and is resume-safe (re-running skips finished videos).

License

MIT. Renderer/physics adapted from jimchen2/pendulum-cjepa-160f (whose own code credits Huawei Technologies Co., Ltd., also MIT).

cat > /tmp/make_pendulum_shards.py <<'PY'
from pathlib import Path
import json
import webdataset as wds

data = Path("/home/jichen/Downloads/pendulum_output/pendulum-cjepa-3k-new")
expected_counts = {"train": 2400, "val": 600}

for split, expected in expected_counts.items():
    source = data / "videos" / split
    files = sorted(source.glob("*.mp4"))
    if len(files) != expected:
        raise RuntimeError(f"{source}: expected {expected} MP4s, found {len(files)}")

    output = data / "wds" / split
    output.mkdir(parents=True, exist_ok=True)
    pattern = str(output / f"pendulum-{split}-%06d.tar")

    with wds.ShardWriter(pattern, maxcount=128) as sink:
        for path in files:
            sink.write({
                "__key__": path.stem,
                "video.mp4": path.read_bytes(),
                "meta.json": json.dumps({
                    "split": split,
                    "filename": path.name,
                }).encode("utf-8"),
            })

    print(f"{split}: wrote {len(files)} videos to {output}")
PY

python /tmp/make_pendulum_shards.py