What the fuck is this model about?
DogOnKeyboard/Mistral-7B-Heretic-V2 is a 7 billion parameter instruct-tuned model based on mistralai/Mistral-7B-Instruct-v0.3. Its primary distinction is that it has been decensored using the Heretic v1.0.1 process.
What makes THIS different from all the other models?
This model stands out due to its significantly reduced refusal rate. While the original Mistral-7B-Instruct-v0.3 refused 86 out of 100 prompts, this Heretic version refuses only 2 out of 100. This decensoring was achieved with a minimal impact on the model's original distribution, as indicated by a low KL divergence of 0.08. It retains the base model's enhancements, including an extended vocabulary (32768), support for a v3 tokenizer, and function calling capabilities.
Should I use this for my use case?
Good for:
- Applications requiring a less restrictive or "decensored" language model.
- Use cases where the original Mistral-7B-Instruct-v0.3's high refusal rate was a limitation.
- Scenarios benefiting from the Mistral architecture's efficiency and performance.
- Developers needing function calling support within a 7B parameter model.
Key Capabilities:
- Decensored Output: Drastically reduced content refusals compared to its base model.
- Instruct Following: Capable of following instructions for various tasks.
- Function Calling: Supports tool use and function calling, enabling integration with external systems.
- Extended Vocabulary: Benefits from the
Mistral-7B-v0.3's larger vocabulary and v3 tokenizer.