finetuning

To fine-tune a model on Azure AI Foundry, you submit training jobs using supervised, preference, or reinforcement learning techniques. This process requires preparing your dataset, configuring the training job, deploying the model, and evaluating the results.

Supervised fine-tuning (SFT) uses labeled datasets for tasks like classification, preference fine-tuning (DPO) optimizes models based on human feedback for ranking, and reinforcement fine-tuning (RFT) trains models with external tool calls for complex decision-making.

Yes, Azure AI Foundry access is required to run reinforcement fine-tuning. You also need compatible Python libraries like azure-identity, azure-ai-projects, and openai to successfully execute the training and deployment workflows.

Yes, you can use preference fine-tuning (DPO) to optimize models based on human feedback. This approach is specifically useful for improving ranking, search, and recommendation systems by aligning model outputs with user preferences.

The azure-identity, azure-ai-projects, and openai Python libraries are required to submit a fine-tuning job. These dependencies enable authentication, project management, and model interaction within the Azure AI Foundry environment.

Supervised fine-tuning (SFT) is the best way to customize a pre-trained model for classification. By providing a labeled dataset, you train the model for specific tasks like information retrieval and image classification.