This repository contains modeling code for the Gazelle Joint Speech Language Model.

For some more details, read our blog post or update.

This inference code is heavily derived from Huggingface's Llava implementation. It is not particularly optimized, but we welcome community contributions.

Checkpoints:

• v0.2: huggingface
• v0.2-dpo: huggingface
• v0.1: huggingface

Links:

• Join us in Discord.
• Original blog post
• v0.1 release notes

We are backproppin' on a budget and these initial checkpoints are not very robust to many real-world considerations. We make no representation of warranty or liability.

The v0.2 model responds to jailbreaks and adversarial attacks. We do not suggest using this in production.

This modeling code is licensed under Apache 2.0. The v0.2 release is derived from Mistral 7B and is released under the same Apache 2.0 license.

The v0.1 pretrained checkpoints derive from Llama 2, which is governed by the Llama 2 license. You must agree to these terms if you use the pretrained weights!