Upload gpt-physics
Browse files- README.md +46 -0
- best_model.pt +3 -0
- checkpoint_epoch0_step500.pt +3 -0
- checkpoint_latest.pt +3 -0
README.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
library_name: pytorch
|
| 6 |
+
tags:
|
| 7 |
+
- physics
|
| 8 |
+
- next-frame-prediction
|
| 9 |
+
- gpt
|
| 10 |
+
- mup
|
| 11 |
+
- rigid-body-dynamics
|
| 12 |
+
- icml-2026
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# gpt-physics
|
| 16 |
+
|
| 17 |
+
A small GPT trained from scratch to predict 2D rigid body physics trajectories. Part of an ICML-2026 study on whether language models can learn physical dynamics from text-encoded simulation data.
|
| 18 |
+
|
| 19 |
+
## Model details
|
| 20 |
+
|
| 21 |
+
- **Architecture**: 6-layer GPT, learned positional embeddings, tied LM head
|
| 22 |
+
- **Tokenizer**: digit-level `PhysicsTokenizer` (custom)
|
| 23 |
+
- **Scaling**: muP for hyperparameter transfer
|
| 24 |
+
- **Training**: curriculum learning over 5 difficulty stages
|
| 25 |
+
- **Task**: autoregressive next-frame prediction over 200-frame rigid-body scenes
|
| 26 |
+
- **Domain**: 2D rigid body dynamics simulated with Pymunk / Chipmunk2D
|
| 27 |
+
|
| 28 |
+
## Files
|
| 29 |
+
|
| 30 |
+
- `best_model.pt` — best validation checkpoint (~69 MB)
|
| 31 |
+
- `checkpoint_latest.pt` — latest training step (~158 MB)
|
| 32 |
+
- `checkpoint_epoch0_step500.pt` — early checkpoint (~158 MB)
|
| 33 |
+
|
| 34 |
+
State dicts contain raw `transformer.*` and `lm_head.*` keys for a stock 6-layer GPT — load with the project's `src/scratch/gpt.py` model class.
|
| 35 |
+
|
| 36 |
+
## Training data
|
| 37 |
+
|
| 38 |
+
Trained on ~900K scenes across 24 "seen" scenario types (collisions, stacking, ramps, constraints, mini-games, complex). See [physics-scenarios-packed](https://huggingface.co/datasets/AlexWortega/physics-scenarios-packed) and [physics-scenarios-raw](https://huggingface.co/datasets/AlexWortega/physics-scenarios-raw).
|
| 39 |
+
|
| 40 |
+
## Intended use
|
| 41 |
+
|
| 42 |
+
Research on whether autoregressive LMs can internalize physical dynamics. Not intended for production physics simulation — use Pymunk for that.
|
| 43 |
+
|
| 44 |
+
## Citation
|
| 45 |
+
|
| 46 |
+
ICML-2026 submission (in progress).
|
best_model.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:986d77152b04a8cbbeab3062bffae51219b5d2df8c9f510a225f2225d05ea212
|
| 3 |
+
size 69389410
|
checkpoint_epoch0_step500.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b0f3e26125fc581e0366dc9842a5d00bcad86df315d2a8d84fa1aa0d35f01e9c
|
| 3 |
+
size 157848890
|
checkpoint_latest.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7ecb31e6d101d01dd412094d5bf479572332867361231ed35737c186db9b60ad
|
| 3 |
+
size 157847482
|