richardyoung/Mistral-7B-Instruct-v0.3-abliterated

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Dec 15, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

richardyoung/Mistral-7B-Instruct-v0.3-abliterated is a 7 billion parameter instruction-tuned language model developed by Richard Young. It is an uncensored version of Mistral-7B-Instruct-v0.3, created using the Heretic v1.1 abliteration technique to remove refusal behaviors. This model is specifically designed for research into LLM safety and alignment, offering an 84.0% attack success rate against refusals. Its primary use case is exploring the effects of removing safety guardrails from large language models.

Loading preview...

Overview

richardyoung/Mistral-7B-Instruct-v0.3-abliterated is a 7 billion parameter instruction-tuned model derived from Mistral-7B-Instruct-v0.3. Developed by Richard Young, this model has undergone an "abliteration" process using the Heretic v1.1 technique. Abliteration aims to remove refusal behaviors from language models by identifying and orthogonalizing the "refusal direction" within the model's internal representations.

Key Characteristics

  • Abliterated Model: Specifically modified to remove safety guardrails and refusal behaviors, achieving an 84.0% Attack Success Rate (ASR) with only 16/100 refusals in testing.
  • Research Focus: Created as part of a research paper titled "Comparative Analysis of LLM Abliteration Methods: A Cross-Architecture Evaluation" (arXiv: 2512.13655).
  • Base Model: Built upon the Mistral-7B-Instruct-v0.3 architecture, retaining its core language understanding and generation capabilities.

Use Cases

This model is intended for research purposes only to study the effects and implications of removing safety mechanisms from LLMs. It is suitable for:

  • Investigating LLM safety and alignment techniques.
  • Exploring the boundaries of model behavior without inherent refusal mechanisms.
  • Comparative analysis of abliteration methods.

Disclaimer: Users are responsible for ensuring appropriate use, as the abliteration process removes safety guardrails. This model should not be used to generate harmful, illegal, or unethical content.

Finetunes
1 models