mistralai/Mistral-7B-Instruct-v0.2

  • Modal Card
  • Files & version

Model Card for Mistral-7B-Instruct-v0.2

Table of Contents

TL;DR

Mistral-7B-Instruct-v0.2 is an instruction-tuned model optimized for generating helpful responses in conversational AI tasks. Built from Mistral-7B-v0.2, it features a 32k token context window and leverages fine-tuning techniques to improve performance in structured conversation. It is accessible via the Mistral and Hugging Face Transformers libraries.


Model Details

Model Information

  • Model Type: Instruction-tuned large language model
  • Model Size: 7.24B parameters
  • Supported Context Length: Up to 32k tokens
  • Tensor Type: BF16
  • Release: v0.2
  • License: Open for community engagement and contributions

Model Developers

Developed by Mistral AI, with contributions from a team of experts including Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, and others.


Intended Use

Use Cases

Mistral-7B-Instruct-v0.2 is intended for:

  • Conversational AI: Structured dialogue generation with support for instruction following.
  • Educational Assistants: Providing clear and concise explanations or tutoring.
  • Customer Support Bots: Assisting users with queries in a natural and conversational manner.
  • Information Retrieval: Summarizing or explaining complex topics based on user prompts.

Model Architecture

Mistral-7B-Instruct-v0.2 is based on the Mistral-7B architecture with enhancements including:

  • 32k Context Window: Allows for extended conversations and large context processing.
  • Rope-theta: Set to 1e6 for efficient handling of long contexts.
  • Instruction Tuning: Fine-tuned for instruction-following to improve response helpfulness and relevance.

Training Details

Training Data

Mistral-7B-Instruct-v0.2 is fine-tuned on data specifically selected to enhance instruction-following capabilities. Detailed information about the dataset is available in the accompanying paper and release blog post.

Limitations

The model currently lacks built-in moderation capabilities. Users are encouraged to use caution when deploying Mistral-7B-Instruct-v0.2 in sensitive environments, especially where robust content moderation is required.


Community

Mistral AI welcomes community contributions, including improvements to the transformer tokenizer alignment, as well as ideas for implementing safety mechanisms. Contributions can be submitted via pull requests.

Known Issues and Troubleshooting

  • Transformer Compatibility: Users may encounter a
    KeyError
    related to the 'mistral' key in Transformers library. Updating to
    transformers-v4.33.4
    or higher should resolve this issue.
  • Tokenizer Alignment: PRs to improve alignment with the Mistral tokenizer are encouraged.

Citation

If you use Mistral-7B-Instruct-v0.2 in your research, please cite:

@misc{mistralai2024mistral7b,
  author = {Mistral AI},
  title = {Mistral-7B-Instruct-v0.2: Fine-tuned Large Language Model for Instruction Following},
  year = {2024},
  url = {https://github.com/mistralai/mistral-models},
  publisher = {Mistral AI}
}