mattshumer/Reflection-Llama-3.1-70B

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Sep 5, 2024License:llama3.1Architecture:Transformer1.7K Featherless Exclusive Warm

Reflection Llama-3.1 70B is a 70 billion parameter open-source LLM, developed by mattshumer, based on Llama 3.1 Instruct. It is trained with a novel Reflection-Tuning technique that enables the model to detect and correct errors in its reasoning process. This model excels at complex reasoning tasks by separating internal thought processes from final outputs, making it suitable for applications requiring high accuracy and self-correction.

Loading preview...

Reflection Llama-3.1 70B Overview

Reflection Llama-3.1 70B is a 70 billion parameter open-source large language model, fine-tuned from Llama 3.1 70B Instruct. Its core innovation is Reflection-Tuning, a technique that teaches the model to identify and rectify mistakes in its reasoning. This training uses synthetic data generated by Glaive.

Key Capabilities & Features

  • Self-Correction: The model can detect errors in its reasoning and attempt to correct them, indicated by <reflection> tags within its thought process.
  • Separated Reasoning: It outputs internal reasoning within <thinking> tags and the final answer within <output> tags, improving clarity and user experience.
  • Standard Llama 3.1 Compatibility: Uses the stock Llama 3.1 chat template and can be sampled with existing Llama code and pipelines.
  • Custom System Prompt: Optimized for use with a specific system prompt that emphasizes complex reasoning and reflection.

Recommended Usage

  • Complex Reasoning Tasks: Ideal for applications where accurate, self-corrected reasoning is critical.
  • Enhanced Output Clarity: Benefits use cases where distinguishing between internal thought processes and final answers is valuable.
  • System Prompt Integration: Best performance is achieved by using the recommended system prompt, potentially combined with custom instructions.
  • Sampling Parameters: Initial recommendations include a temperature of .7 and top_p of .95 for optimal results. Appending "Think carefully." to messages can also increase accuracy.