gpt-oss: How to Run & Fine-tune
OpenAI releases 'gpt-oss-120b' and 'gpt-oss-20b', two SOTA open language models under the Apache 2.0 license. Both 128k context models outperform similarly sized open models in reasoning, tool use, and agentic tasks. You can now run & fine-tune them locally with Unsloth!
Trained with RL, gpt-oss-120b rivals o4-mini and gpt-oss-20b rivals o3-mini. Both excel at function calling and CoT reasoning, surpassing o1 and GPT-4o.
OpenAI released a standalone parsing and tokenization library called Harmony which allows one to tokenize conversations to OpenAI's preferred format for gpt-oss. The official OpenAI cookbook article provides many more details on how to use the Harmony library.
Inference engines generally use the jinja chat template instead and not the Harmony package, and we found some issues with them after comparing with Harmony directly. If you see below, the top is the correct rendered form as from Harmony. The below is the one rendered by the current jinja chat template. There are quite a few differences!
We also made some functions to directly allow you to use OpenAI's Harmony library directly without a jinja chat template if you desire - you can simply parse in normal conversations like below:
