physics-scenarios-raw

Raw (un-tarred) JSONL version of the 2D rigid body physics dataset. Each scene is a separate file under <split>/<scenario_type>/scene_<id>.jsonl. Streaming-friendly for HF datasets and curriculum sampling.

For the bandwidth-efficient packaged version, see physics-scenarios-packed.

Scale (this snapshot)

• Train: 80,000 scenes
• Val: 10,000 scenes
• Test: 10,000 scenes
• Frames per scene: 200
• Format: one .jsonl per scene (1 header + 200 frame lines)

This is a subset of the full 1M-scene dataset; the full dataset is in physics-scenarios-packed.

Layout

train/
  000000/scene_000000.jsonl
  000000/scene_000001.jsonl
  ...
  000079/...
val/
  ...
test/
  ...

Files are sharded into directories of ~1000 scenes (zero-padded shard ids) to keep listings tractable.

Schema (per scene file)

Line 1 — scene header:

{
  "scenario": "billiards",
  "difficulty": 3,
  "static_geometry": [...],
  "constraints": [...],
  "objects": [{"id": 0, "type": "circle", "radius": 0.5, "mass": 1.0, ...}, ...]
}

Lines 2–201 — per-frame state:

{"frame": 0, "objects": [{"id": 0, "x": 1.234, "y": 5.678, "vx": ..., "vy": ..., "angle": ..., "omega": ...}, ...]}

Loading

from datasets import load_dataset
ds = load_dataset("AlexWortega/physics-scenarios-raw", split="train", streaming=True)

Or stream directly from the hub with huggingface_hub.HfFileSystem.

Citation

ICML-2026 submission (in progress).