Hey there! If you’re interested in fine-tuning GPT-OSS models, you’re in luck. Unsloth now offers a way to fine-tune GPT-OSS-20B for free on Colab. This is a game-changer, as most training methods and libraries require at least 40GB VRAM. But Unsloth has managed to fit it into just 14GB VRAM.
The team has also identified and fixed some issues with differing implementations of the GPT-OSS model, which can affect inference performance. These include extra newlines in Jinja chat templates, incorrect rendering of tool calling, missing `<|channel|>final` tags, and numerical instability when running in float16 machines.
To get started, you can check out Unsloth’s Colab notebook for fine-tuning GPT-OSS-20B. They’ve also provided linearized versions of the model for fine-tuning and post-training purposes. And, if you have questions, they’ve got you covered with some Frequently Asked Questions.
What do you think about this development? Are you excited to try fine-tuning GPT-OSS for free?