Model Card for rustral-2-2k-2024-1-6

This is an instruction finetune of Mistral-7B-v0.1 on a custom dataset of Rust code snippets. The current version was finetuned on 2k examples for 3 epochs, with instructions being extracted from the comments of the code itself. Future versions will be based on more data and generated instructions.

Model Details

Model Description

Developed by: Leon Luithlen
Model type: Mistral-7B-v0.1 variant
Language(s) (NLP): Rust, English
License: apache-2.0
Finetuned from model [optional]: Mistral-7B-v0.1

Uses

This model is intended as a code assistant that develops short snippets, such as a single Rust struct or function, from the immediately preceding code and an instruction.

Direct Use

After compiling to the appropriate quantization size for your hardware with llama.cpp, it can be used like other Mistral-7B variants with the following input format:

        Below is an instruction that describes a task, paired with an input that
        provides further context. Write a response that appropriately completes the
        request.

        ### Instruction:
        {instruction}

        ### Input:
        {up_to_two_previous_snippets}

        ### Response:

Out-of-Scope Use

Any programming language other than Rust, natural languages

Bias, Risks, and Limitations

Currently not evaluated systematically, as there is no agreed benchmark for code assistance tasks in Rust.

For uses outside its intended scope, the model is likely to behave similarly to Mistral-7B, but probably worse.

Recommendations

The same caveats that apply to Mistral-7B apply to this model as well.

Training Details

Training Data

Training data is sourced from open source and permissively licensed Rust projects, but is not public.