| sae: | |
| type: matryoshka_batch_topk | |
| activation_dim: 4096 | |
| expansion_factor: 4 | |
| layer_id: 15 | |
| hookpoint: residual | |
| k: 256 | |
| group_fractions: | |
| - 0.5 | |
| - 0.25 | |
| - 0.125 | |
| - 0.0625 | |
| - 0.0625 | |
| group_weights: null | |
| trainer: | |
| epochs: 1 | |
| lr: null | |
| l1_penalty: 0.1 | |
| warmup_steps: 10 | |
| sparsity_warmup_steps: 0 | |
| decay_start: null | |
| resample_steps: null | |
| seed: 42 | |
| device: cuda:0 | |
| log_every_n_steps: 1000 | |
| logger_type: mlflow | |
| validate: false | |
| auxk_alpha: 0.03125 | |
| threshold_beta: 0.999 | |
| threshold_start_step: 1000 | |
| threshold_dead_features: 100000 | |
| data: | |
| dataset_names: | |
| - mimic_findings_temporal | |
| activations_type: per_token | |
| num_workers: 18 | |
| batch_size: 8192 | |
| val_samples: 512000 | |
| train_samples: null | |
| norm_act: true | |
| norm_to_sqrt_act_dim: false | |
| input_unit_norm: false | |
| filter_dict: null | |