zaakirio/LFM2.5-1.2B-Instruct-Uncensored
zaakirio/LFM2.5-1.2B-Instruct-Uncensored is a 1.2 billion parameter LFM2 architecture model, derived from LiquidAI/LFM2.5-1.2B-Instruct. It has been uncensored using the Heretic tool's directional ablation technique, significantly reducing refusal rates from 98% to 5% on harmful prompts while maintaining a low KL divergence of 0.1003 on harmless prompts. This model is designed for research and unrestricted local use where the original model's safety alignment is not desired, offering a 32768 token context length.
Loading preview...
LFM2.5-1.2B-Instruct-Uncensored Overview
This model, developed by zaakirio, is an uncensored variant of the LiquidAI/LFM2.5-1.2B-Instruct base model. It leverages the Heretic tool's directional ablation technique to remove safety alignment, drastically reducing refusal rates without extensive fine-tuning or human prompt engineering. The model maintains a 32768 token context length and is provided in full-precision BF16 format.
Key Capabilities
- Significantly Reduced Refusals: Achieves a refusal rate of just 5% on harmful prompts, down from 98% in the original model.
- Preserved Original Behavior: 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.
- Efficient Uncensoring: Utilizes an automated TPE optimizer to select ablation parameters, ensuring an optimal trade-off between refusal reduction and behavioral preservation.
- Direct Load: Provided as a merged, full-precision BF16 model, requiring no adapter merge or dequantization steps.
Good For
- Research into Model Alignment: Ideal for studying the effects of safety alignment removal and model behavior without censorship.
- Unrestricted Local Use: Suitable for applications where the base model's safety guardrails are not desired or are managed externally.
- Exploring Model Limitations: Useful for understanding the inherent capabilities and biases of the base model when unconstrained by safety filters.