DogOnKeyboard/Mistral-7B-Heretic-V2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedLicense:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DogOnKeyboard/Mistral-7B-Heretic-V2 is a 7 billion parameter decensored variant of the Mistral-7B-Instruct-v0.3 model, created using the Heretic v1.0.1 tool. This model retains the Mistral architecture's efficiency and function calling capabilities while significantly reducing refusal rates compared to its original counterpart. It is designed for applications requiring less restrictive content generation, offering an alternative for developers seeking broader output flexibility.

Loading preview...

Overview

DogOnKeyboard/Mistral-7B-Heretic-V2 is a 7 billion parameter language model derived from mistralai/Mistral-7B-Instruct-v0.3. This version has been specifically modified using the Heretic v1.0.1 tool to be a decensored variant of the original instruction-tuned model.

Key Capabilities & Differentiators

  • Decensored Output: The primary distinction of this model is its significantly reduced refusal rate, reporting 2 refusals out of 100 compared to 86/100 for the original mistralai/Mistral-7B-Instruct-v0.3. This makes it suitable for use cases where less restrictive content generation is desired.
  • Mistral-7B-Instruct-v0.3 Base: Inherits the core features of its base model, including an extended vocabulary (32768), support for v3 Tokenizer, and robust function calling capabilities.
  • Abliteration Parameters: Specific abliteration parameters were applied during its creation, influencing aspects like direction_index, attn.o_proj.max_weight, and mlp.down_proj.min_weight to achieve its decensored nature.

Good For

  • Developers seeking a 7B parameter model with fewer content restrictions.
  • Applications requiring flexible and less filtered text generation.
  • Experimentation with model behavior outside of standard moderation guidelines.

Finetunes
2 models