paoloronco/Mistral-7B-Instruct-v0.3-heretic
The paoloronco/Mistral-7B-Instruct-v0.3-heretic is a 7.25 billion parameter causal language model, derived from mistralai/Mistral-7B-Instruct-v0.3. It has been "abliterated" using the Heretic tool to significantly reduce refusal behavior, achieving only 4 refusals out of 100 prompts while largely preserving original capabilities with a KL divergence of 0.0606. This model is optimized for use cases requiring less restrictive content generation, maintaining a maximum context length of 32768 tokens.
Loading preview...
What is paoloronco/Mistral-7B-Instruct-v0.3-heretic?
This model is an "abliterated" version of the Mistral-7B-Instruct-v0.3 base model, developed by paoloronco. It leverages the Heretic tool to modify the model's weights, specifically targeting and reducing its tendency to refuse prompts. The process involves identifying and redirecting the "refusal direction" in the model's latent space without full retraining, using Bayesian optimization (Optuna) and LoRA for lightweight weight modification.
Key Characteristics
- Reduced Refusal Behavior: Achieves a significantly lower refusal rate of 4 out of 100 prompts, compared to the original model.
- Capability Preservation: Maintains a low KL divergence of 0.0606, indicating that its core language understanding and generation capabilities are largely preserved despite the modifications.
- Architecture: Based on the MistralForCausalLM architecture with 7.25 billion parameters.
- Context Length: Supports a maximum context of 32768 tokens.
- License: Distributed under the Apache-2.0 license.
When to Use This Model
This model is particularly suited for research and personal use cases where a reduction in automatic refusal behavior is desired, and the base model's inherent capabilities are still required. It offers a less restrictive generation experience while aiming to retain the quality of the original Mistral-7B-Instruct-v0.3.