Reflection Llama 3.1 70B

🚀 Reflection Llama-3.1 70B

Reflection Llama-3.1 70B is an open - source LLM. It's trained with Reflection - Tuning, a new technique enabling the LLM to detect and correct reasoning mistakes.

⚠️ Important Note

There was an issue with the model when we first uploaded it. If you tried it and didn't have good results, please try again. We think we've fixed the issue.

The model was trained on synthetic data generated by Glaive. If you're training a model, Glaive is highly recommended.

You can try the model here.

✨ Features

• Reflection - Tuning Technique: Trained with Reflection - Tuning, which allows the model to detect and correct reasoning errors.
• Separation of Thoughts and Answers: The model can separate internal thoughts and reasoning from the final answer using special tags like <thinking> and <output>.
• Error Detection and Correction: Can detect mistakes in reasoning and correct them within <reflection> tags.

📋 Model Information

📚 Documentation

Benchmarks

Trained from Llama 3.1 70B Instruct, you can sample from Reflection Llama - 3.1 70B using the same code, pipelines, etc. as any other Llama model. It uses the stock Llama 3.1 chat template format, with a few new special tokens trained for reasoning and reflection.

During sampling, the model starts by outputting reasoning inside <thinking> and </thinking> tags. Once satisfied with its reasoning, it outputs the final answer inside <output> and </output> tags. These tags are special tokens trained into the model.

The <thinking> section may contain one or more <reflection> tags, indicating that the model has detected an error in its reasoning and will correct it before giving the final answer.

System Prompt

The system prompt used for training this model is:

You are a world - class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.

We recommend using this exact system prompt to get the best results from Reflection Llama - 3.1 70B. You can also experiment by combining it with your own custom instructions to customize the model's behavior.

Chat Format

The model uses the standard Llama 3.1 chat format. Here's an example:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a world - class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.<|eot_id|><|start_header_id|>user<|end_header_id|>

what is 2+2?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Tips for Performance

💡 Usage Tip

• We initially recommend a temperature of .7 and a top_p of .95.
• For increased accuracy, append Think carefully. at the end of your messages.

Dataset / Report

Both the dataset and a brief report detailing how we trained this model will be released next week. Alongside, our Reflection 405B model is expected to be the top - performing LLM in the world, including closed - source models.

Thanks to Jason Kuperberg and Josh Bickett from the HyperWrite team for reviewing drafts of the report we'll be releasing next week.

Also, we know right now the model is split into a ton of files. We'll condense this soon to make the model easier to download and work with!