If you’re building .NET apps that rely on fine-tuned models, this month’s Foundry updates are worth paying attention to. Reinforcement Fine-Tuning just got more accessible and significantly cheaper.
The full details are in the official announcement, but here’s the practical breakdown.
Global Training for o4-mini
o4-mini is the go-to model for reasoning-heavy and agentic workloads. The big news: you can now launch fine-tuning jobs from 13+ Azure regions with lower per-token training rates compared to Standard training. Same infrastructure, same quality, broader reach.
If your team is spread across geographies, this matters. You’re no longer pinned to a handful of regions to train.
Here’s the REST API call to kick off a global training job:
curl -X POST "https://<your-resource>.openai.azure.com/openai/fine_tuning/jobs?api-version=2025-04-01-preview" \
-H "Content-Type: application/json" \
-H "api-key: $AZURE_OPENAI_API_KEY" \
-d '{
"model": "o4-mini",
"training_file": "<your-training-file-id>",
"method": {
"type": "reinforcement",
"reinforcement": {
"grader": {
"type": "string_check",
"name": "answer-check",
"input": "{{sample.output_text}}",
"reference": "{{item.reference_answer}}",
"operation": "eq"
}
}
},
"hyperparameters": {
"n_epochs": 2,
"compute_multiplier": 1.0
},
"trainingType": "globalstandard"
}'
That trainingType: globalstandard flag is the key difference.
New Model Graders: GPT-4.1 Family
Graders define the reward signal your model optimizes against. Until now, model-based graders were limited to a smaller set of models. Now you get three new options: GPT-4.1, GPT-4.1-mini, and GPT-4.1-nano.
When should you reach for model graders instead of deterministic ones? When your task output is open-ended, when you need partial credit scoring across multiple dimensions, or when you’re building agentic workflows where tool-call correctness depends on semantic context.
Here’s the thing – the tiering strategy is practical:
- GPT-4.1-nano for initial iterations. Low cost, fast feedback loops.
- GPT-4.1-mini once your grading rubric is stable and you need higher fidelity.
- GPT-4.1 for production grading or complex rubrics where every scoring decision counts.
You can even mix grader types in a single RFT job. Use string-match for the “correct answer” dimension and a model grader for evaluating reasoning quality. That flexibility is honestly what makes this useful for real workloads.
The RFT Data Format Gotcha
This trips people up. RFT data format is different from SFT. The last message in each row must be a User or Developer role – not Assistant. The expected answer goes in a top-level key like reference_answer that the grader references directly.
If you’ve been doing supervised fine-tuning and want to switch to RFT, you need to restructure your training data. Don’t skip this step or your jobs will fail silently.
Why This Matters for .NET Developers
If you’re calling fine-tuned models from your .NET apps through the Azure OpenAI SDK, cheaper training means you can iterate more aggressively. The model grader options mean you can fine-tune for nuanced tasks – not just exact-match scenarios. And the best practices guide on GitHub will save you real debugging time.
Start small. Ten to a hundred samples. Simple grader. Validate the loop. Then scale.
