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 directional ablation using the 'heretic' method to minimize refusal rates while preserving original model behavior, making it an 'uncensored' variant. It maintains the original chat template and tool-calling tokens, offering a model with reduced safety alignments for specific applications. The model has a context length of 32768 tokens.
Loading preview...
Model Overview
alpharomercoma/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter language model based on LiquidAI's LFM2.5-1.2B-Instruct. This version has been specifically modified to remove refusal behaviors, making it an "abliterated" or "uncensored" model. The modification was performed using the heretic method, which involves directional ablation of the refusal direction within the residual stream. This process uses a TPE search to optimize per-layer ablation weights, aiming to minimize both the refusal rate and the KL divergence from the original model.
Key Characteristics
- Abliterated Refusals: The primary differentiator is the intentional removal of safety-related refusal mechanisms present in the base model.
- Preserved Core Functionality: Despite the ablation, the model retains the original chat template and tool-calling tokens from the LiquidAI base model.
- Merged Weights: The modifications are merged directly into the model weights, allowing it to load as a standard
Lfm2ForCausalLM. - Context Length: Supports a substantial context window of 32768 tokens.
- ExecuTorch Export: An ExecuTorch program (
.pte) is available for CPU deployment, optimized for XNNPACK backend with 8-bit dynamic activations and 4-bit grouped weights.
Use Cases
This model is suitable for applications where the removal of inherent safety training and refusal behaviors is explicitly desired. Users should exercise responsibility, as the safety training of the base model has been deliberately removed.