richardyoung/LFM2.5-1.2B-Instruct-heretic
richardyoung/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter instruction-tuned language model, derived from LiquidAI's LFM2.5-1.2B-Instruct, with a 32,768 token context length. This version has been 'abliterated' using the Heretic v2.0.0.dev0 tool to significantly reduce safety refusals while maintaining original capabilities. It features a hybrid convolution + attention architecture optimized for fast on-device inference and is suitable for research, creative writing, and unrestricted local inference.
Loading preview...
Model Overview
This model, richardyoung/LFM2.5-1.2B-Instruct-heretic, is a 1.2 billion parameter instruction-tuned language model based on LiquidAI's LFM2.5-1.2B-Instruct. It has been modified using the Heretic v2.0.0.dev0 tool through a process called "abliteration." This technique removes the "refusal direction" from the model's residual stream, effectively disabling safety refusals without retraining the entire model.
Key Differentiators
- Decensored Output: Achieves a 97% reduction in refusals compared to the original model (3/100 vs. 98/100 refusals).
- Capability Retention: Maintains original model capabilities and coherence with a very low KL divergence of 0.059, well below the 0.5 "damage" threshold.
- On-Device Optimization: Inherits the LFM2.5 architecture, designed for fast on-device deployment with a hybrid convolution + attention architecture.
- Efficient Inference: Supports various inference frameworks including Ollama,
llama.cpp,llama-cpp-python, and Hugging Face Transformers.
Intended Use Cases
- Research: Particularly in alignment techniques and model behavior.
- Creative Writing: For generating diverse and unrestricted content.
- Education: For exploring and understanding model safety mechanisms.
- Unrestricted Local Inference: For applications requiring responses without built-in safety refusals, used responsibly and in accordance with applicable laws.