Fine-Tuning in Foundry

The Fine-Tuning Process

Supported Models for Fine-Tuning

Global Training (2026 Feature)

As of April 2026, Foundry supports Global Training for models like o4-mini. This allows you to launch fine-tuning jobs across 13+ Azure regions, offering lower per-token training rates compared to standard regional training.

Reinforcement Fine-Tuning (RFT)

For reasoning models (o-series), Foundry provides Reinforcement Fine-Tuning (RFT). Unlike Supervised Fine-Tuning (which teaches formatting or style), RFT aligns model behavior with complex business logic by explicitly rewarding accurate reasoning paths.

Training Data Format (SFT JSONL)

{"messages": [
  {"role": "system", "content": "You are a legal contract analyzer."},
  {"role": "user", "content": "Analyze this NDA clause: ..."},
  {"role": "assistant", "content": "Risk Level: Medium. Key concerns: ..."}
]}

Fine-Tuning Costs

• Training — Charged per token processed during training
• Hosting — Hourly fee while the model is deployed (even when idle)
• Inference — Per-token, typically higher than base models