openchat

Why we included this project

OpenChat is a family of instruction-tuned models built on Llama-2 and Llama-3, and the 7B and 8B releases are small enough to run on a single consumer GPU like an RTX 3090. The more interesting piece is the training method: C-RLFT, an approach inspired by offline reinforcement learning that handles mixed-quality data without preference labels. That part is worth reading even if you plan to fine-tune your own models instead of just downloading the weights. The repo also includes the training and evaluation scripts used to produce and benchmark each release, and the Apache-2.0 license keeps commercial use straightforward. For teams that want a self-hosted assistant without a big hardware budget, it works as both a deployable model and a clear reference for how it was built.