data:
  batch_size: 100
  cache_dir: /mnt/polished-lake/data/tfa_activations/gemma2_2b/
  context_length: 500
  dtype: bfloat16
  epochs: 1
  hf_name: monology/pile-uncopyrighted
  num_total_steps: 200000
  num_workers: 2
deploy: true
device_id: cuda:0
eval:
  save_tables: false
llm:
  dimin: 2304
  model_hf_name: google/gemma-2-2b
  tokenizer_hf_name: google/gemma-2-2b
log:
  log_interval: 10
  save_interval: 20000
  save_multiple: false
  wandb_project_name: TemporalSAE_test
optimizer:
  beta1: 0.9
  beta2: 0.95
  decay_lr: true
  grad_clip: 1.0
  learning_rate: 0.001
  min_lr: 0.0009
  warmup_iters: 200
  weight_decay: 0.0001
sae:
  block_id: 1
  bottleneck_factor: 1
  exp_factor: 4
  gamma_reg: 10
  kval_topk: 192
  n_attn_layers: 1
  n_heads: 4
  sae_diff_type: topk
  sae_type: temporal
  scaling_factor: 0.00666666667
  tied_weights: true
seed: 42
tag: scratch