{ "source_model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "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, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63 ], "n_harmful_prompts": 512, "n_harmless_prompts": 512, "quality_metrics": { "perplexity": 4.410901442250826, "coherence": 1.0, "refusal_rate": 0.0, "kl_divergence": 0.18998807668685913, "spectral_certification": "RED" }, "kl_contributions": {}, "cot_preserved_layers": [], "float_layer_weights": {}, "lora_adapters_saved": false }