Spaces:
Running
Running
refactor: implement centralized configuration, upgrade SAE training to multi-layer TopK, and optimize dashboard attribution UX
b7ddfc6 | model: | |
| n_layers: 2 | |
| n_heads: 4 | |
| d_model: 128 | |
| max_length: 30 | |
| data: | |
| env_id: "MiniGrid-Empty-8x8-v0" | |
| num_episodes: 1000 | |
| collection_method: "PPO-Teacher" | |
| interpretability: | |
| dla_threshold: 0.1 | |
| patching_metric: "logit_diff" | |
| sae: | |
| expansion_factor: 8 | |
| l1_coeff: 0.0005 | |
| num_episodes: 100 | |