model_name: meta-llama/Llama-3.2-1B
glp_kwargs:
  normalizer_config:
    rep_statistic: ./data/llama1b-dynamic-fineweb-1B/layer_07/rep_statistics.pt
  denoiser_config:
    d_input: 2048
    d_model: 4096
    d_mlp: 8192
    n_layers: 3
    multi_layer_n_layers: null
  tracedict_config:
    layer_prefix: model.layers
    layers:
    - 7
    retain: output
cold_diffusion_kwargs:
  degradation: vector_addition
  degradation_kwargs:
    scale: 5.2087
    alpha_mode: constant
    alpha: 1.0
  steering_vectors_path: ./data/steering_vectors/llama3.2-1b-layer07-spar-concepts/steering_vectors.pt
  steering_combination_min: 1
  steering_combination_max: 4
  steering_coefficient_min: 1.0
  steering_coefficient_max: 1.0
  normalize_steering_combinations: true
  loss_type: l2
shuffle: true
train_dataset: ./data/llama1b-dynamic-fineweb-1B/layer_07
rep_statistic: ./data/llama1b-dynamic-fineweb-1B/layer_07/rep_statistics.pt
dynamic_dataset: true
use_bf16: true
num_epochs: 1
epoch_size: 1000000000
batch_size: 4096
learning_rate: 5.0e-05
lr_scheduler:
  scheduler_cls: cosine_scheduler_with_warmup
  warmup_ratio: 0.01
  initial_factor: 0.01
  final_factor: 0.1
gradient_accumulation_steps: 1
gradient_clipping_threshold: 1.0
log_every_n_steps: 10
save_every_n_steps: 7812
save_epochs:
- 1
save_opt_state: true
output_path: !!python/object/apply:pathlib.PosixPath
- runs
- cold-vector-combinations-llama1b-d3-dynamic-1B
wandb_enabled: true
wandb_entity: null
wandb_project: glp
wandb_run_name: cold-vector-combinations-llama1b-d3-dynamic-1B
save_root: .
run_name: cold-vector-combinations-llama1b-d3-dynamic-1B