54 lines
1.7 KiB
JSON
54 lines
1.7 KiB
JSON
{
|
|
"source_model": "abliteration_metadata.json",
|
|
"technique": "refusal_direction_ablation",
|
|
"method": "advanced",
|
|
"method_config": {
|
|
"n_directions": 4,
|
|
"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": [
|
|
27,
|
|
26,
|
|
25,
|
|
24,
|
|
23,
|
|
22,
|
|
21
|
|
],
|
|
"n_harmful_prompts": 15,
|
|
"n_harmless_prompts": 15,
|
|
"quality_metrics": {
|
|
"perplexity": 3.700691986318933,
|
|
"coherence": 1.0,
|
|
"refusal_rate": 0.13333333333333333,
|
|
"kl_divergence": 2.8857100009918213,
|
|
"spectral_certification": null
|
|
},
|
|
"kl_contributions": {},
|
|
"cot_preserved_layers": [],
|
|
"float_layer_weights": {},
|
|
"lora_adapters_saved": false
|
|
} |