ai/mistral repository overview

⁠Mistral 7B Instruct v0.2

A fast and powerful 7B parameter model excelling in reasoning, code, and math. Mistral 7B is a powerful 7.3B parameter language model that outperforms Llama 2 13B across a wide range of benchmarks, including reasoning, reading comprehension, and code generation. Despite its smaller size, it delivers performance comparable to much larger models, making it efficient and versatile.

⁠Intended uses

Mistral 7B is designed to provide high-quality responses across a wide range of general-purpose NLP tasks while remaining efficient in resource usage. Also, this model is fine-tuned to follow instructions, allowing it to perform tasks and answer questions naturally. The base model doesn’t have this capability.

• Automated code generation: Automates creation of code snippets, reducing manual coding and accelerating development.
• Debugging support: Identifies code errors and provides actionable recommendations to streamline debugging.
• Text summarization and classification: Supports summarizing text, classification, and text/code completion tasks.
• Conversational applications: Fine-tuned for conversational interactions using diverse datasets.
• Knowledge retrieval: Delivers accurate, detailed answers for enhanced information retrieval.
• Mathematical accuracy: Reliably processes and solves complex mathematical problems.
• Roleplay and text generation: Generates extensive narrative text for roleplaying and creative scenarios.

⁠Characteristics

i: Estimated

⁠Available model variants

¹: VRAM estimated based on model characteristics.

⁠Use this AI model with Docker Model Runner

First, pull the model:

docker model pull ai/mistral

Then run the model:

docker model run ai/mistral

For more information on Docker Model Runner, explore the documentation⁠.

⁠Considerations

• Best suited for English.
• Performs well out-of-the-box but can be fine-tuned further.
• Use appropriate system prompts for safer and more controlled outputs.
• To use instruction fine-tuning, wrap your prompt with [INST] and [/INST] tags. The first instruction must start with a beginning-of-sentence token, while any following instructions should not. The assistant's response will automatically end with an end-of-sentence token.

⁠Benchmark performance

⁠Links

• Mistral 7b⁠
• Mistral 7b-Paper⁠