mattshumer/ref_70_e3
mattshumer/ref_70_e3 is a 70 billion parameter Llama 3.1-based instruction-tuned large language model developed by mattshumer. It is trained with a novel Reflection-Tuning technique, enabling it to detect and correct errors in its reasoning process. This model is designed to separate internal thought processes from final outputs, improving the reliability and clarity of its responses. It is particularly suited for complex reasoning tasks where self-correction is beneficial.
Loading preview...
Reflection Llama-3.1 70B Overview
mattshumer/ref_70_e3 is a 70 billion parameter open-source large language model based on Llama 3.1 Instruct. Its key differentiator is a novel Reflection-Tuning technique, which trains the model to identify and rectify mistakes in its own reasoning. This process involves generating synthetic data via Glaive to teach self-correction.
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 and Output: It outputs internal reasoning within
<thinking>and</thinking>tags, and the final answer within<output>and</output>tags, enhancing clarity and user experience. - Llama 3.1 Compatibility: Uses the standard Llama 3.1 chat template and can be sampled with existing Llama code and pipelines.
- Custom System Prompt: Optimized for a specific system prompt that emphasizes complex reasoning and reflection, allowing for custom instruction combinations.
Recommended Usage
To achieve optimal performance, users should employ the recommended system prompt and consider appending "Think carefully." to messages for increased accuracy. The model is particularly well-suited for applications requiring robust reasoning and transparent thought processes, where the ability to self-correct can significantly improve outcome quality.