saidutta69/DeepSeek-R1-Distill-Llama-8B-heretic
The saidutta69/DeepSeek-R1-Distill-Llama-8B-heretic is an 8 billion parameter language model, a decensored variant of DeepSeek-R1-Distill-Llama-8B. Developed by RACER IS OP using Heretic v1.4.0, it suppresses refusal behavior through targeted weight edits rather than fine-tuning, preserving the base model's strong reasoning and instruction-following capabilities. This model is designed for developers seeking a capable 8B reasoning model that provides direct answers without built-in refusal guardrails.
Loading preview...
Overview
This model, created by RACER IS OP, is a decensored variant of the deepseek-ai/DeepSeek-R1-Distill-Llama-8B model. It utilizes Heretic v1.4.0 (directional ablation or "abliteration") to suppress refusal behavior. Unlike traditional fine-tuning, abliteration directly edits specific weights responsible for refusal, leaving the base model's knowledge and instruction-following largely intact. This approach aims to avoid the coherence degradation often seen when fine-tuning a "helpful" persona on top of RLHF'd refusals.
Key Capabilities
- Decensored Responses: Suppresses refusal behavior, allowing the model to answer requests that the base model would typically decline.
- Preserved Reasoning: Maintains the strong reasoning capabilities inherited from the DeepSeek-R1-Distill-Llama-8B base model.
- Instruction Following: Keeps the original model's instruction-following abilities largely intact due to the targeted weight editing method.
- Hardware Friendly: Can be run on consumer hardware with 8-12 GB GPUs or via GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0).
Good For
- Developers who require a capable 8B reasoning model that provides direct answers without built-in safety filtering or refusal guardrails.
- Use cases where the base model's knowledge and instruction-following are critical, and censorship is undesirable.
- Experimentation with models that have had refusal behaviors removed via targeted weight edits.