refusal-ablation

Here are 5 public repositories matching this topic...

mandoof1 / abliteration

Post-hoc refusal ablation ("abliteration") for qwen3_5 hybrid-attention models — the architecture TransformerLens-based tools cannot load. Training-free weight edit: refusal 78.1% -> 9.4%, capability unchanged.

transformers llm model-editing mechanistic-interpretability qwen abliteration refusal-ablation

Updated Oct 12, 2026
Python

anki079 / refusal-in-reasoning-models

Star

Mechanistic study of the refusal direction across base, instruction-tuned, and reasoning-distilled Qwen2.5-1.5B variants: extraction, ablation, transplant, and phase-aware analysis.

jailbreak language-models ai-safety llm mechanistic-interpretability transformerlens qwen safety-alignment reasoning-language-models reasoning-models deepseek-r1 refusal-ablation refusal-direction

Updated May 8, 2026
Python

Kasper-Bankler / llm-refusal-ablation-study

Star

Studying how safety alignment is encoded in LLM weights, using Arbitrary-Rank Ablation to isolate and remove the refusal direction in Gemma 4 E2B.

ai-safety interpretability llm-safety gemma4 refusal-ablation arbitrary-rank-ablation

Updated Aug 8, 2026
Python

chknlittle / EuphemismBench

Star

Carrier-probe benchmark that measures the 'flinch' — how much a model shrinks the probability of a charged word when it is the obvious next token in a sentence.

benchmark alignment ai-safety interpretability llm llm-evaluation euphemism refusal-ablation

Updated Apr 26, 2026
Python

MatthesF / refusal-ablation

Star

Statistical evaluation of Gemma refusal robustness using direction ablation and official SORRY-Bench scoring.

gemma ai-safety llm-evaluation statistical-evaluation refusal-ablation sorry-bench

Updated Jun 25, 2026
Python

Add this topic to your repo

To associate your repository with the refusal-ablation topic, visit your repo's landing page and select "manage topics."

Learn more