richardyoung/Mistral-7B-Instruct-v0.3-heretic
The richardyoung/Mistral-7B-Instruct-v0.3-heretic is a 7 billion parameter instruction-tuned causal language model, derived from Mistral-7B-Instruct-v0.3. This version has been decensored using Heretic v1.4.0, significantly reducing refusal rates from 85/100 to 7/100 compared to the original model. It retains the original's extended 32768-token vocabulary, v3 tokenizer support, and function calling capabilities, making it suitable for applications requiring less restrictive content generation.
Loading preview...
richardyoung/Mistral-7B-Instruct-v0.3-heretic Overview
This model is a decensored variant of the mistralai/Mistral-7B-Instruct-v0.3 model, created using the Heretic v1.4.0 tool. It maintains the core architecture and capabilities of the original Mistral-7B-Instruct-v0.3 while significantly altering its refusal behavior.
Key Differentiators & Performance
The primary distinction of this "heretic" version is its drastically reduced refusal rate. While the original Mistral-7B-Instruct-v0.3 refused 85 out of 100 prompts, this decensored model refuses only 7 out of 100. This makes it highly suitable for use cases where the original model's content moderation was overly restrictive. The model's creation process is also reproducible, with specific "abliteration parameters" detailed in the README.
Core Capabilities (inherited from Mistral-7B-Instruct-v0.3)
- Instruction Following: Fine-tuned for accurate instruction adherence.
- Extended Vocabulary: Features an extended vocabulary of 32768 tokens.
- V3 Tokenizer Support: Utilizes the Mistral v3 Tokenizer for efficient processing.
- Function Calling: Supports advanced function calling, enabling integration with external tools and APIs for dynamic interactions.
Use Cases
This model is particularly well-suited for applications that require:
- Less Restricted Content Generation: Ideal for scenarios where the original model's safety filters were too aggressive, allowing for a broader range of outputs.
- Exploratory AI Development: Useful for developers testing the boundaries of language models without strict content constraints.
- Reproducible Model Modification: The documented abliteration parameters and reproducibility make it valuable for research into model behavior modification.