Justbackup/LFM2.5-1.2B-Instruct-Uncensored
Justbackup/LFM2.5-1.2B-Instruct-Uncensored is a 1.2 billion parameter LFM2 architecture model, derived from LiquidAI's LFM2.5-1.2B-Instruct. This version has been uncensored using directional ablation with the Heretic tool, significantly reducing refusal rates while maintaining original model behavior. It is optimized for research and local use cases requiring a less restrictive instruction-following model.
Loading preview...
Model Overview
Justbackup/LFM2.5-1.2B-Instruct-Uncensored is a 1.2 billion parameter instruction-tuned model based on LiquidAI's LFM2.5-1.2B-Instruct. Its primary differentiator is the removal of safety alignment, or "censorship," achieved through directional ablation using the Heretic tool. This process automatically optimizes to minimize refusal rates while preserving the original model's behavior as much as possible, without human fine-tuning.
Key Features & Performance
- Uncensored Behavior: Achieves a significantly lower refusal rate of 5% on harmful prompts, compared to 98% for the original model.
- Behavioral Fidelity: Maintains a low KL divergence of 0.1003 on harmless prompts, indicating that its responses on benign inputs remain very close to the original model.
- Architecture: Built on the LFM2 architecture with 16 layers, operating in BF16 precision.
- Deployment Ready: Provided as a merged, full-precision BF16 model, requiring no adapter merge or dequantization steps at load time.
Intended Use
This model is designed for research and unrestricted local use where the removal of refusal behavior is desired. Users are responsible for its deployment and ensuring compliance with all applicable laws and the base model's license.