alpharomercoma/LFM2.5-1.2B-Instruct-heretic
alpharomercoma/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter instruction-tuned causal language model, derived from LiquidAI/LFM2.5-1.2B-Instruct. This model has undergone a directional ablation process using 'heretic' to minimize refusal rates while preserving original model performance, making it an "uncensored" variant. It maintains the base model's chat template and tool-calling capabilities, optimized for use cases requiring less restrictive content generation.
Loading preview...
Model Overview
alpharomercoma/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter instruction-tuned language model, specifically an "abliterated" or "uncensored" version of the original LiquidAI/LFM2.5-1.2B-Instruct. This modification was achieved using the heretic method, which performs directional ablation of the refusal direction in the residual stream.
Key Differentiators
- "Uncensored" Output: The primary distinction is the intentional removal of the base model's safety training, aiming to reduce refusal rates in content generation. This was done through a targeted ablation process that minimizes both refusal and KL divergence from the original model.
- Performance Preservation: The
hereticprocess involved a TPE search over per-layer ablation weights to ensure that while refusal rates are minimized, the model's overall performance and characteristics (like chat template and tool-calling tokens) remain largely unchanged from the base model. - ExecuTorch Support: An ExecuTorch program (
lfm2_5_1_2b_heretic_8da4w.pte) is provided, optimized for CPU inference with 8-bit dynamic activations and 4-bit grouped weights, leveraging XNNPACK and KleidiAI kernels on Arm hosts.
Use Cases
This model is suitable for applications where less restrictive content generation is desired, particularly in scenarios where the base model's safety filters might be overly cautious. Developers should use it responsibly, acknowledging the deliberate removal of safety training.