| { | |
| "trainer": { | |
| "trainer_class": "BSFTrainer", | |
| "dict_class": "BlockSparseSAE", | |
| "variant": "vanilla_block_topk", | |
| "lr": 5e-05, | |
| "grad_clip": 1.0, | |
| "snr": 0.1, | |
| "auxk_alpha": 0.0, | |
| "k_anneal_steps": 0, | |
| "top_k_aux_blocks": 32, | |
| "steps": 50000, | |
| "warmup_steps": 1000, | |
| "decay_start": 40000, | |
| "resample_interval": 1000, | |
| "resample_dead_steps": 400, | |
| "resample_boost": 1.5, | |
| "resample_max_per_round": 512, | |
| "selection": "bias_topk", | |
| "l1_coeff": 0.003, | |
| "l1_warmup_steps": 5000, | |
| "bias_gamma": 0.002, | |
| "bias_clamp": 1.0, | |
| "bias_boost_below": 0.0001, | |
| "bias_nuke_above": 0.5, | |
| "k": 16, | |
| "dict_size": 2048, | |
| "code_dim": 131072, | |
| "activation_dim": 4096, | |
| "block_sizes": { | |
| "64": 2048 | |
| }, | |
| "act_scale": 262.5943908691406, | |
| "seed": 0, | |
| "device": "cuda", | |
| "layer": 24, | |
| "lm_name": "Qwen/Qwen3-8B", | |
| "wandb_name": "bsf_stream_L24_G2048_b64_k16_biasdz0.0001-0.5_trainer_0", | |
| "submodule_name": "resid_post_layer_24" | |
| } | |
| } |