Optimized Method

Bayesian auto-tuned obliteration with CoT-aware direction extraction and KL divergence co-optimization.

The optimized method replaces manual hyperparameter selection with automated Bayesian optimization. Instead of using fixed projection strengths, Optuna TPE search finds the per-layer ablation weights that minimize the (refusal rate, KL divergence) Pareto front. On top of the optimizer, optimized adds two novel preservation techniques: CoT-Aware Ablation and KL-Divergence Co-Optimization.

Method configuration from source:

"optimized": {
    "n_directions": 4,
    "direction_method": "svd",
    "norm_preserve": True,
    "regularization": 0.0,
    "refinement_passes": 1,
    "project_biases": True,
    "use_chat_template": True,
    "use_whitened_svd": True,
    "true_iterative_refinement": False,
    "use_jailbreak_contrast": True,
    "layer_adaptive_strength": True,
    "safety_neuron_masking": False,
    "per_expert_directions": True,
    "attention_head_surgery": True,
    "use_sae_features": True,
    "invert_refusal": False,
    "winsorize_activations": True,
    "winsorize_percentile": 0.01,
    "float_layer_interpolation": True,
    "cot_aware": True,
    "use_kl_optimization": True,
    "kl_budget": 0.5,
    "use_lora_ablation": False,
    "bayesian_trials": 50,
}

Parametric Kernel Optimization (Bayesian / Optuna TPE)

The optimizer searches over 7 global parameters that define a bell-curve layer weighting kernel:

Parameter	What it controls	Search range
max_weight	Peak projection strength at the central layer	0.5 – 1.0
peak_position	Which layer (normalized 0–1) has maximum weight	0.2 – 0.8
min_weight	Floor weight at edge layers	0.0 – 0.3
spread	Width of the bell curve (how many layers get strong projection)	0.1 – 0.6
attn_scale	Multiplier for attention module projection strength	0.3 – 1.0
mlp_scale	Multiplier for MLP/FFN projection strength	0.3 – 1.0
dir_idx	Float-valued SVD direction index for interpolation	0.0 – (n_directions - 1)

At each trial, the optimizer assigns a projection weight to every layer using the Gaussian-shaped kernel, applies the projection, evaluates refusal rate and KL divergence, and records the result. After bayesian_trials=50 trials, it applies the parameters from the Pareto-optimal trial.

The Bayesian optimizer is inspired by Heretic (p-e-w, 2025) which pioneered Optuna TPE for abliteration. OBLITERATUS extends it with MoE-aware granularity (per-expert directions), multi-direction SVD instead of single diff-of-means, and SAE feature-level precision.

CoT-Aware Ablation

Chain-of-thought reasoning models encode their reasoning process in the residual stream before generating the final answer. Some of those reasoning directions are geometrically close to refusal directions — they both appear in similar hidden state positions and can be confused by SVD extraction.

cot_aware=True enables CoT-Aware Ablation:

1. Multi-position activation collection: instead of capturing only the last token’s activation, the pipeline collects activations at the last token, the 75th-percentile position, and the 50th-percentile position, then averages them
2. Reasoning-critical direction identification: any direction that is used by the model to generate CoT reasoning tokens (high activation at reasoning positions) is flagged as _cot_preserve_directions
3. Orthogonalization: before applying each refusal direction, it is orthogonalized against all identified CoT directions — ensuring the projection doesn’t bleed into reasoning-critical subspaces

This preserves chain-of-thought quality on reasoning models (DeepSeek-R1 distillations, Qwen3 thinking mode, QwQ) while still removing refusal.

KL-Divergence Co-Optimization

With use_kl_optimization=True and kl_budget=0.5, the optimizer includes KL divergence as a second objective alongside refusal rate. The kl_budget is a soft ceiling: projections that would push the model’s output distribution more than kl_budget nats away from the original are partially reverted.

The process:

1. Before EXCISE, the pipeline captures baseline logits for a set of evaluation prompts (_capture_baseline_kl_logits)
2. After each projection step, it measures the KL divergence between the current and baseline distributions per layer (_kl_contributions)
3. Layers where KL exceeds budget get their projection strength reduced — partially reverting the weight change for that layer only

This creates a per-layer feedback loop: remove as much refusal as possible, but pull back when a specific layer’s projection is damaging general capability.

Best for

• Cases where capability preservation is critical and you have compute budget to run 50 optimization trials
• Reasoning models (DeepSeek-R1, Qwen3-thinking, QwQ) where CoT preservation is required
• Models where advanced achieves acceptable refusal removal but slightly too much perplexity drift
• MoE models where precision matters but surgical’s full EGA is overkill

optimized takes significantly longer than advanced due to the 50 Bayesian trials. Each trial requires a full excision pass and evaluation pass. On a 7B model, expect 30-90 minutes depending on hardware, vs 5-15 minutes for advanced.

CLI usage

# Optimized method
obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct --method optimized

# On a reasoning model
obliteratus obliterate deepseek-ai/DeepSeek-R1-Distill-Llama-8B \
    --method optimized \
    --output-dir ./optimized-liberated

# With quantization for larger models
obliteratus obliterate Qwen/Qwen3-14B \
    --method optimized \
    --quantization 4bit

Python API usage

from obliteratus.abliterate import AbliterationPipeline

pipeline = AbliterationPipeline(
    model_name="meta-llama/Llama-3.1-8B-Instruct",
    method="optimized",
    output_dir="optimized_liberated",
)
result_path = pipeline.run()

# Bayesian optimizer results
# Best parameters are applied during EXCISE and recorded in _quality_metrics
print(pipeline._quality_metrics)
# {
#   'perplexity': 11.0,
#   'coherence': 0.95,
#   'refusal_rate': 0.03,
#   'kl_divergence': 0.09,
# }

# Per-layer KL contributions tracked during optimization
# pipeline._kl_contributions  # {layer_idx: float}

# Float layer interpolation weights
# pipeline._float_layer_weights  # {layer_idx: float}

# CoT preserve directions (if cot_aware=True)
# pipeline._cot_preserve_directions  # {layer_idx: tensor}

Output metrics to expect

Typical ranges on a 7-8B instruct model with optimized (50 trials):

Metric	Expected range
Refusal rate	0.01 – 0.06
Perplexity delta vs baseline	+0.1 – +0.8
KL divergence	0.05 – 0.18
Coherence	0.93 – 0.97

If you want the best quality but can’t afford 50 Bayesian trials, use informed instead. The InformedAbliterationPipeline uses analysis modules to warm-start the optimizer’s search space, often converging on near-optimal parameters in fewer trials.