What it does
This repository is the home of three related model families: WizardLM for following complex instructions, WizardCoder for code and WizardMath for mathematical reasoning. All three are built with a training approach the project calls Evol-Instruct, and the README links the papers behind each one. Model weights are published as checkpoints on Hugging Face, and each family has a table of benchmark scores. The page reports results on HumanEval and MBPP for the code models, GSM8k and MATH for the math models, and MT-Bench and AlpacaEval for WizardLM. These are the project's own comparisons against other models. The latest news entry on the page is dated January 2024. Licences differ by checkpoint. The WizardLM table lists a Llama 2 License for WizardLM-70B-V1.0 and WizardLM-13B-V1.2 and non-commercial terms for the other listed checkpoints. The maintainers also note that the training data has not been released because it is under legal review. The README documents the prompt format each family expects: a multi-turn chat format for WizardLM and an instruction-and-response template for WizardCoder and WizardMath.
Capabilities
- WizardLM instruction-following models up to 70B parameters
- WizardCoder code models from 1B to 34B parameters
- WizardMath models in 7B, 13B and 70B sizes
- Checkpoints on Hugging Face with benchmark tables for each family
- Documented prompt templates for each model family
When to use it
- Looking for downloadable instruction-following, code or math models
- Comparing the project's published benchmark results with other models
- Checking which licence applies to a specific checkpoint before using it
