richardyoung/Llama-3.1-8B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jun 24, 2026License:llama3.1Architecture:Transformer0.0K Featherless Exclusive Cold

richardyoung/Llama-3.1-8B-Instruct-heretic is an 8 billion parameter instruction-tuned causal language model, a decensored version of Meta's Llama-3.1-8B-Instruct, created using the Heretic v1.4.0 tool. This model is specifically modified to reduce refusals, demonstrating 3 refusals out of 100 compared to the original's 96/100. It maintains the Llama 3.1 architecture with an 8192 token context length and is optimized for multilingual dialogue use cases, excelling in scenarios where reduced content moderation is desired.

Loading preview...

Overview

This model, richardyoung/Llama-3.1-8B-Instruct-heretic, is an 8 billion parameter instruction-tuned variant of Meta's Llama-3.1-8B-Instruct. It was created using the Heretic v1.4.0 tool, specifically designed to produce a "decensored" version of the original model. The primary differentiator is its significantly reduced refusal rate, with only 3 refusals out of 100 compared to the original model's 96/100, as measured by the Heretic project.

Key Capabilities

  • Decensored Responses: Engineered to provide responses with a substantially lower refusal rate than the base Llama 3.1 Instruct model.
  • Llama 3.1 Foundation: Inherits the robust capabilities of the Meta Llama 3.1 architecture, including strong performance in general reasoning, code generation, and mathematical tasks.
  • Multilingual Support: Optimized for multilingual dialogue, supporting languages such as English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
  • Tool Use: Supports advanced tool use and function calling, with detailed guidance available in the Llama 3.1 prompt format documentation.

Good For

  • Use cases requiring less restrictive content moderation or "decensored" outputs.
  • Applications where the original Llama 3.1 Instruct model's safety alignments are too conservative.
  • Research into model safety and alignment, particularly in understanding the impact of decensoring techniques.
  • Multilingual assistant-like chat applications where a broader range of responses is desired.

Finetunes
1 models