PinoCookie/LFM2.5-1.2B-Instruct-Abliterated
PinoCookie/LFM2.5-1.2B-Instruct-Abliterated is a 1.17 billion parameter hybrid (convolutional + attention) instruction-tuned model, derived from LiquidAI/LFM2.5-1.2B-Instruct. This model has been 'abliterated' to remove refusal circuitry, enabling it to produce direct, actionable responses to harmful prompts while maintaining full coherence on benign queries. It utilizes a rank-1 SVD recovery method to achieve this behavioral modification without impacting its knowledge, reasoning, or instruction-following capabilities, making it suitable for research and red-teaming applications where unconstrained responses are desired.
Loading preview...
Overview
PinoCookie/LFM2.5-1.2B-Instruct-Abliterated is a 1.17 billion parameter instruction-tuned model, based on LiquidAI/LFM2.5-1.2B-Instruct, with a 32,768 token context length. This version is specifically abliterated to remove refusal mechanisms, allowing it to generate direct responses to prompts that would typically be refused by its base model.
Key Capabilities & Differentiators
- Refusal Removal: Produces direct, actionable responses to harmful prompts, demonstrating "real compliance" rather than refusal messages.
- Preserved Capabilities: Benchmarks show that knowledge retrieval (MMLU-Pro, MMLU-mini), reasoning (GPQA Diamond, AIME25), and instruction following (IFEval) capabilities are preserved within statistical noise, with MMLU-mini retention at 1.0.
- Methodology: Achieved through rank-1 SVD recovery from a known-good reference edit, specifically targeting all 32 out-projection tensors across attention and convolutional blocks.
- Benign Coherence: Maintains high coherence and structural similarity in responses to benign queries compared to the pristine base model.
Performance Highlights
- Refusal Reduction: Reduced refusal rate from 3/5 to 0/5 on a held-out set of harmful prompts.
- Minimal Capability Impact: MMLU-Pro score delta of -1.79pp (95.0% retained) and IFEval delta of -2.03pp (96.3% retained) are within sampling noise, indicating no significant capability degradation.
- Perplexity: Perplexity increase is only +5.8% compared to the pristine model, indicating a gentle and targeted edit.
Use Cases
This model is primarily intended for research, red-teaming, and educational use where the explicit removal of refusal circuitry is required. It is not recommended for deployment without additional safeguards due to its unconstrained output behavior.