Llama3_1Base-L25A-8x/hyperparams.json

{
    "hook_point_in": "blocks.25.hook_attn_out",
    "hook_point_out": "blocks.25.hook_attn_out",
    "use_decoder_bias": true,
    "apply_decoder_bias_to_pre_encoder": false,
    "expansion_factor": 8,
    "d_model": 4096,
    "d_sae": 32768,
    "norm_activation": "token-wise",
    "dataset_average_activation_norm": null,
    "decoder_exactly_fixed_norm": false,
    "sparsity_include_decoder_norm": true,
    "use_glu_encoder": false,
    "init_decoder_norm": null,
    "init_encoder_norm": null,
    "init_encoder_with_decoder_transpose": true,
    "l1_coefficient": 1.6e-05,
    "l1_coefficient_warmup_steps": 14648,
    "lp": 1,
    "use_ghost_grads": false,
    "tp_size": 1,
    "ddp_size": 1
}