PCBWorld / README.md
hyungseoksongh's picture
Model card: policy architecture
0445eb7 verified
|
Raw History Blame Contribute Delete
1.66 kB
---
library_name: pytorch
license: other
license_name: exaone
license_link: https://huggingface.co/LG-AI-Research/PCBWorld/blob/main/LICENSE
tags:
- reinforcement-learning
- pcb-routing
- kicad
- pcbworld
---
# PCBWorld checkpoints
Trained routing policies from the PCBWorld paper. Weights only. All are decoder-only transformer policies
(4 layers, d_model 128, 8 heads, FFN 512; 1.6M parameters) trained from scratch, seeds 42–45.
- Paper: [PCBWorld: A Benchmark Environment for Engine-Grounded PCB Design Automation](https://arxiv.org/abs/2607.05915)
- Code: [github.com/LGAI-Research/PCBWorld](https://github.com/LGAI-Research/PCBWorld)
| Folder | Dataset | Agent |
|---|---|---|
| `d1/grid{10,50,100,200,500}/` | D1-10 … D1-500 (synthetic grid-based boards) | PPO |
| `d2/ppo/` | D2 (synthetic gridless boards) | PPO |
| `d2/ppo_terminal/` | D2 | PPO (terminal) |
| `d2/grpo/` | D2 | GRPO |
| `d2/ppo_no_finish/` | D2 | PPO (w/o finish) |
| `d2/reward_ablation/wire<λw>_via<λv>/` | D2 | PPO, wirelength weight λw × via weight λv |
`d2/reward_ablation/wire0.002_via0.1/` is the default weighting and holds the same files as `d2/ppo/`.
Each file holds the weights (`policy_state_dict`) and the training arguments (`args`).
```bash
hf download LG-AI-Research/PCBWorld --local-dir var/checkpoints/pcbworld
```
## License
EXAONE AI Model License Agreement 1.2 - NC ([LICENSE](https://huggingface.co/LG-AI-Research/PCBWorld/blob/main/LICENSE)): research and educational use only.
Commercial use of the model and of its output, including boards routed with it, is not permitted.
The PCBWorld code is licensed separately; see its repository.