AlexWortega commited on
Commit
f1ae877
·
verified ·
1 Parent(s): 14a69b2

Add README

Browse files
Files changed (1) hide show
  1. README.md +92 -0
README.md ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - text-generation
7
+ tags:
8
+ - physics
9
+ - rigid-body
10
+ - pymunk
11
+ - next-frame-prediction
12
+ - simulation
13
+ - icml-2026
14
+ size_categories:
15
+ - 1M<n<10M
16
+ ---
17
+
18
+ # physics-scenarios-packed
19
+
20
+ Packed (tar.gz) version of a 2D rigid body physics dataset for training language models on next-frame prediction. **1,000,020 scenes × 200 frames** simulated with Pymunk / Chipmunk2D.
21
+
22
+ This repo is bandwidth-friendly: each scenario type ships as a single `.tar.gz`. For the unpacked JSONL files see [physics-scenarios-raw](https://huggingface.co/datasets/AlexWortega/physics-scenarios-raw).
23
+
24
+ ## Scale
25
+
26
+ - **Train**: 900,000 scenes (24 seen scenario types × 37,500 each)
27
+ - **Val**: 100,020 scenes (30 scenario types × 3,334 each — includes 6 OOD types)
28
+ - **Frames per scene**: 200
29
+ - **Format**: JSONL inside tar.gz (1 header line + 200 frame lines per scene file)
30
+
31
+ ## Layout
32
+
33
+ ```
34
+ manifest.json # full dataset spec
35
+ train/
36
+ avalanche.tar.gz # 24 seen-type archives
37
+ basketball.tar.gz
38
+ ...
39
+ val/
40
+ avalanche.tar.gz # 30 archives (24 seen + 6 OOD)
41
+ pong.tar.gz # OOD types — never seen at train time
42
+ bowling.tar.gz
43
+ ...
44
+ ```
45
+
46
+ ## Scenario types
47
+
48
+ **Seen (24)** — train + val: avalanche, basketball, billiards, breakout, bridge, chain, conveyor, dominos, explosion, funnel, head_on, jenga, marble_run, orbit, pendulum, pinball, plinko, projectile, pyramid, seesaw, ski_jump, tower, wind, wrecking_ball
49
+
50
+ **Unseen / OOD (6)** — val only:
51
+ - *Simple*: pong, bowling, ramp_roll
52
+ - *Complex*: angry_birds, hourglass, newtons_cradle
53
+
54
+ The 6 unseen types are held out to measure generalization to scenarios with new dynamics or interactions.
55
+
56
+ ## Schema (per scene file)
57
+
58
+ Line 1 — header:
59
+ ```json
60
+ {"scenario": "billiards", "difficulty": 3, "static_geometry": [...], "constraints": [...], "objects": [...]}
61
+ ```
62
+
63
+ Lines 2–201 — frames:
64
+ ```json
65
+ {"frame": 0, "objects": [{"id": 0, "x": 1.234, "y": 5.678, "vx": ..., "vy": ..., "angle": ..., "omega": ...}, ...]}
66
+ ```
67
+
68
+ ## Categories
69
+
70
+ The 30 scenario types span 6 categories:
71
+ - **Collision** — head_on, billiards, breakout, bowling, pong
72
+ - **Stacking** — jenga, tower, pyramid, dominos
73
+ - **Ramp** — ski_jump, projectile, ramp_roll, marble_run
74
+ - **Constraint** — pendulum, chain, bridge, newtons_cradle, hourglass
75
+ - **Minigame** — pinball, plinko, basketball, angry_birds
76
+ - **Complex** — avalanche, explosion, wind, wrecking_ball, conveyor, funnel, orbit, seesaw
77
+
78
+ ## Usage
79
+
80
+ ```python
81
+ from huggingface_hub import snapshot_download
82
+ import tarfile
83
+
84
+ snapshot_download("AlexWortega/physics-scenarios-packed", repo_type="dataset", local_dir="data")
85
+
86
+ with tarfile.open("data/train/billiards.tar.gz") as tar:
87
+ tar.extractall("data/train_extracted")
88
+ ```
89
+
90
+ ## Citation
91
+
92
+ ICML-2026 submission (in progress). Generation pipeline: `src/physics/scenario_generator.py`.