Together AI Enhances Fine-Tuning API with Long-Context and Conversational Data Support

In a strategic move to bolster its Fine-Tuning API, Together AI has announced a suite of new features aimed at enhancing the customization and efficiency of large language models (LLMs). As organizations increasingly leverage generative AI for competitive advantages, the ability to fine-tune these models for specific tasks becomes paramount, according to Together AI.

Longer-Context Fine-Tuning

The Fine-Tuning API now supports longer-context fine-tuning, which is crucial for handling large documents and complex data inputs. This feature extends the context window up to 32K for the Llama 3.1 8B and 70B models. This advancement is particularly beneficial for tasks requiring extensive content retention and interpretation, such as document review and long-form content generation.

Conversational and Instruction Data Format Support

With the rise of applications like chatbots and virtual assistants, Together AI has introduced support for conversational and instruction data formats. This allows developers to directly input conversation histories into the API, streamlining the preparation process and eliminating the need for manual data formatting. This feature is compatible with the OpenAI fine-tuning API, facilitating a seamless transition for developers using existing datasets.

Training Quality Improvements

Together AI has implemented several enhancements to its training processes, resulting in more capable models without increasing costs. These improvements have been validated through rigorous testing on mathematics and knowledge-based benchmarks, showing significant performance boosts across various tasks.

Validation Dataset Support

The introduction of validation dataset support allows users to monitor model performance on unseen data during training. This feature aids in optimizing hyperparameters and refining the training setup, ensuring models can generalize effectively to new examples.

Quality-of-Life Enhancements

Several additional improvements have been made to enhance user experience and model performance. These include better integration with Weights & Biases for experiment tracking, automated batch size settings for optimal training efficiency, and more flexible learning rate schedules.

Together AI's comprehensive update to its Fine-Tuning API underscores its commitment to providing robust and customizable AI solutions. These enhancements not only improve model quality but also offer users greater control and flexibility in their AI development endeavors.