| { | |
| "trainer": { | |
| "trainer_class": "BatchTopKCrossCoderTrainer", | |
| "dict_class": "BatchTopKCrossCoder", | |
| "lr": 0.0001, | |
| "steps": 97656, | |
| "auxk_alpha": 0.03125, | |
| "warmup_steps": 1000, | |
| "decay_start": null, | |
| "threshold_beta": 0.999, | |
| "threshold_start_step": 1000, | |
| "top_k_aux": 2048, | |
| "seed": null, | |
| "activation_dim": 4096, | |
| "dict_size": 131072, | |
| "k": 1000, | |
| "k_target": 200, | |
| "k_initial": 1000, | |
| "k_annealing_steps": 5000, | |
| "code_normalization": "CROSSCODER", | |
| "code_normalization_alpha_sae": 1.0, | |
| "code_normalization_alpha_cc": 0.1, | |
| "device": "cuda", | |
| "layer": 16, | |
| "lm_name": "meta-llama/Meta-Llama-3.1-8B-Instruct-meta-llama/Meta-Llama-3.1-8B", | |
| "wandb_name": "Meta-Llama-3.1-8B-L16-k200-lr1e-04-local-shuffling-Crosscoder-ni0.3-ka1k5k", | |
| "submodule_name": null, | |
| "dict_class_kwargs": { | |
| "same_init_for_all_layers": "True", | |
| "norm_init_scale": "0.3", | |
| "init_with_transpose": "True", | |
| "encoder_layers": "None", | |
| "code_normalization": "crosscoder", | |
| "code_normalization_alpha_sae": "1.0", | |
| "code_normalization_alpha_cc": "0.1" | |
| } | |
| } | |
| } |