finetuning

To fine-tune Azure AI Foundry models, you prepare and validate JSONL chat datasets, then submit SFT, DPO, or RFT training jobs. The process automates dataset conversion, validation, and job submission while monitoring training outcomes for reliable end-to-end customization.

SFT, DPO, and RFT are distinct fine-tuning techniques supported on Azure AI Foundry. SFT uses supervised labels, DPO leverages preference data, and RFT reinforces reasoning workflows. The process validates JSONL formats and manages job submissions for all three training types.

Yes, you need OpenAI-compatible Azure clients for text and vision fine-tuning on Azure AI Foundry. The workflow requires these clients along with azure-identity and azure-ai-projects dependencies to authenticate, submit jobs, and enforce deployment-specific format and SKU rules.

To evaluate and deploy a fine-tuned Azure AI Foundry model, you analyze training curves, assess held-out test sets, and compare checkpoints against a baseline. The workflow then deploys the best model version and cleans up expired resources after experiments conclude.

To validate a JSONL dataset before submitting DPO or SFT jobs, use deterministic validation scripts to inspect and convert your data. The workflow ensures SFT, DPO, and RFT JSONL datasets meet required format specifications before launching Azure AI Foundry training.

Limitations of SFT for vision fine-tuning workflows include strict JSONL chat format requirements and deployment-specific SKU rules. The workflow manages these constraints by validating datasets, calibrating hyperparameters, and providing fallback paths for supported model families.