WizardChatML-7B-v0

WizardChatML 7B v0

GGUF Quants

I personally think ChatML is the best prompt format. It allows:

1. Easier templating for generation
2. Lower risk of inadvertently generating role tokens
3. Better long-context performance and higher quality on quantized models

This model is an experiment attempting to extend WizardLM 2 7B to ChatML. It was trained on a small ChatML dataset, it probably isn't as good as WizardLM 2 Base, but it's an attempt.

Aside from using the ChatML prompt format, this model supports system prompts. In fact, it adheres very well to these prompts.

If you want to use this model for task-specific purposes, you should probably fine-tune it.

Capabilities & Challenges

• Seems ok-ish at writing
• Pretty good at math
• Sometimes calls itself ChatGPT/OpenAI

Risks

It has not been trained on guardrail data and may generate offensive content if prompted.

License

If you use this model, you must include the Apache 2.0 license AND the following notice:

I'm releasing this model under the Apache 2.0 license, with the added restriction that it cannot be used to compete with OpenAI (due to the nature of the training data). Additionally, this model was finetuned from the WizardLM 2 7B model, which was recently removed by Microsoft (it was Apache licensed, but may have been trained on NC-licensed data). You are responsible for the usage of this model. You are responsible for checking that your usage of this model is legal in your jurisdiction. Commercial use is not advised, as this model is finetuned from a model that may have been trained on NC-licensed data. Make sure to consult a lawyer before using in production or commercially.