ChiKoi7/Llama3.1-SuperHawk-8B-Heretic-v2
ChiKoi7/Llama3.1-SuperHawk-8B-Heretic-v2 is an 8 billion parameter language model, a merge of Llama-3.1-SuperNova-8B-Lite_TIES_with_Base-Heretic and Llama-3.1-Hawkish-8B-Heretic. This model is specifically designed to minimize refusals, achieving 5/100 compared to 99/100 in its original SuperHawk counterpart, by applying the Heretic tool during its creation. It is optimized for use cases where reduced model refusal rates are critical.
Loading preview...
Model Overview
ChiKoi7/Llama3.1-SuperHawk-8B-Heretic-v2 is an 8 billion parameter language model derived from a merge of two 'Heretic-abliterated' models: Llama-3.1-SuperNova-8B-Lite_TIES_with_Base-Heretic and Llama-3.1-Hawkish-8B-Heretic. The core innovation of this model lies in its creation process, where the constituent models were processed with the Heretic tool before merging, specifically to reduce refusal rates.
Key Differentiators
- Significantly Reduced Refusals: This model boasts a refusal rate of 5/100, a substantial improvement over the original Llama3.1-SuperHawk-8B's 99/100 refusal rate. This was achieved by applying the Heretic tool to the base models prior to merging, and then again to the final merged output.
- Controlled Merge Process: The merge utilized the same configuration as the original SuperHawk model, but with the critical step of pre-abliterating the source models using Heretic v1.1.0.
- KL Divergence: The model exhibits a KL divergence of 0.0493, indicating a controlled modification from its base components.
Ideal Use Cases
- Applications requiring a language model with a very low propensity for refusal.
- Scenarios where the original SuperHawk model's high refusal rate was a limiting factor.
- Research into the effects of 'abliteration' tools like Heretic on model behavior and safety alignment.