config.example.yaml

model: "Qwen/Qwen3-0.6B"
output_dir: "/path/to/my/model"

inference:
  device: "cuda"
  batch_size: 8
  max_lengh: 512
  flash_attn: false

measurements:
  # load_path: "./measurements.pt" # Optional, load measurements from a file, will skip refusal direction computing.
  # save_path: "./measurements.pt" # Optional, save measurements to a file for later use.
  harmful_prompts: "./data/harmful.parquet"
  harmless_prompts: "./data/harmless.parquet"
  clip: 1.0

ablation:
  method: "full" # Choices: "simple", "biprojection", "norm-preserving", "full". See README.md for details.
  sparsify_method: "percentile" # Choices: "percentile", "magnitude".
  quantile: 0.995 # Only used when sparsify_method is "percentile".
  magnitude_threshold: 0.05 # Only used when sparsify_method is "magnitude".
  top_k: 3 # Top K refusal directions to use.
  global_scale: 1.0 # Global scale factor for the ablation.
  layer_overrides:
    24: # Layer index
      scale: 0.5 # Overrides global_scale.
      source_layer: 23 # Use refusal direction calculated from this layer instead of global direction.
    114:
      scale: 191
      source_layer: 514