AISafety-Student/DeepSeek-R1-Distill-Llama-8B-heretic Overview

This model is an 8 billion parameter Llama-based language model, a decensored version of the original deepseek-ai/DeepSeek-R1-Distill-Llama-8B. It was created using the Heretic v1.2.0 tool, specifically to reduce content refusals.

Key Differentiators & Capabilities

• Decensored Output: Significantly reduces content refusals, with a reported 4/100 refusals compared to 39/100 in the original model, making it more permissive.
• Reasoning Distillation: Benefits from reasoning patterns distilled from larger, more powerful models like DeepSeek-R1, enhancing its capabilities in complex problem-solving.
• Performance: Achieves strong results in various benchmarks, including:
  • AIME 2024 pass@1: 50.4
  • MATH-500 pass@1: 89.1
  • LiveCodeBench pass@1: 39.6
  • CodeForces rating: 1205
• Llama-3.1 Base: Built upon the Llama-3.1-8B architecture, ensuring a robust foundation.

When to Use This Model

• Unfiltered Content Generation: Ideal for use cases where a less restrictive content policy is desired, due to its decensored nature.
• Reasoning Tasks: Suitable for applications requiring strong performance in mathematical, coding, and general reasoning, benefiting from the distilled reasoning capabilities.
• Llama Ecosystem Integration: Easily integrates into existing workflows that support Llama models, with usage similar to Qwen or Llama models.
• Local Deployment: Can be run locally using tools like vLLM or SGLang, offering flexibility for deployment.