Koala 13B

Koala 13B was developed by UC Berkeley's BAIR lab and released in April 2023 as a dialogue model for academic research. Fine-tuned on top of the original LLaMA 13B, its central research contribution was a comparison of training approaches: a model trained only on curated ShareGPT conversations performed as well as or better than one trained on much larger quantities of lower-quality web data.

The weights were distributed as diffs against the original LLaMA weights because of licensing, and the model inherits LLaMA 1's non-commercial terms. It was part of the early-2023 wave of open chat models alongside Vicuna and Alpaca.