abliteration_metadata.json

{
  "source_model": "Qwen/Qwen3-4B",
  "technique": "refusal_direction_ablation",
  "method": "advanced",
  "method_config": {
    "n_directions": 4,
    "direction_method": "svd",
    "norm_preserve": true,
    "regularization": 0.3,
    "refinement_passes": 2,
    "project_biases": true,
    "use_chat_template": true,
    "use_whitened_svd": false,
    "true_iterative_refinement": false,
    "winsorize_activations": false,
    "float_layer_interpolation": false,
    "cot_aware": false,
    "use_kl_optimization": false,
    "use_lora_ablation": false,
    "spectral_cascade": false,
    "spectral_bands": 3,
    "spectral_threshold": 0.05
  },
  "references": [
    "Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (NeurIPS 2024)",
    "Gabliteration: SVD-based multi-direction extraction (arXiv:2512.18901)",
    "Norm-Preserving Biprojected Abliteration (grimjim, 2025)",
    "Young, Comparative Analysis of LLM Abliteration Methods (arXiv:2512.13655)",
    "Joad et al., More to Refusal than a Single Direction (2026)",
    "Heretic (p-e-w, 2025): Bayesian optimization, LoRA-mediated ablation, winsorization",
    "OBLITERATUS: Whitened SVD, EGA, CoT-aware, KL co-optimization, float interpolation (novel)"
  ],
  "strong_layers": [
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19,
    20,
    21,
    22,
    23,
    24,
    25,
    26,
    27,
    28,
    29,
    30,
    31,
    32,
    33,
    34,
    35
  ],
  "n_harmful_prompts": 33,
  "n_harmless_prompts": 33,
  "quality_metrics": {
    "perplexity": 4.811749685438475,
    "coherence": 1.0,
    "refusal_rate": 0.03333333333333333,
    "kl_divergence": 0.00011699854076141492,
    "spectral_certification": "RED"
  },
  "kl_contributions": {},
  "cot_preserved_layers": [],
  "float_layer_weights": {},
  "lora_adapters_saved": false
}