Together AI Supports Fine-Tuning of Hugging Face Models

Together AI and Hugging Face have collaborated to announce a new feature that allows fine-tuning of any compatible LLM on the Hugging Face Hub through Together AI's infrastructure.

Users perform the task by specifying two models:

• Base Model: An official catalog model from Together AI that serves as a 'training template,' handling GPU resource allocation, memory optimization, training pipelines, and inference configuration.
• Custom Model: The Hugging Face model that the user actually wants to fine-tune.

This integration feature works bidirectionally. Together AI can fetch and train not only public models from Hugging Face but also models from private repositories via an API token. Once training is complete, the model can be automatically uploaded to the Hugging Face Hub through the hf_output_repo_name setting to be shared or deployed.

Currently, all CausalLM models under 100B parameters are designed to be able to use this feature. This allows developers to quickly customize the latest models they discover on Hugging Face to fit their own data and apply them to production environments, without having to build complex training infrastructure.