Read on for the details.<\/p>\n
\n
Global Training for o4-mini<\/h2>\n
Global Training expands the reach of model customization with the affordable pricing of our other Global offerings. With this update, o4-mini<\/strong> joins the list of models you can train globally:<\/p>\n Currently available regions:<\/strong> East US 2, North Central US, West US 3, Australia East, France Central, Germany West Central, Switzerland North, Norway East, Poland Central, Spain Central, Italy North, Switzerland West, and Sweden Central.<\/p>\n o4-mini is one of the most popular models for reasoning-intensive and agentic workloads. Adding Global Training support makes it significantly more cost-effective to customize at scale\u2014especially for teams spread across multiple geographies.<\/p>\n Graders are the engine of RFT\u2014they define the reward signal your model optimizes against. Until now, model-based graders were limited to a smaller set of models. Starting this month, three additional models are available as graders:<\/p>\n Deterministic graders (string-match, Python, endpoint-based) should remain your default\u2014they are faster, cheaper, and more reproducible. Reach for model graders when:<\/p>\n Tip:<\/strong> You can mix grader types within a single RFT job. For example, use a string-match grader for the “correct answer” dimension and a GPT-4.1-mini model grader for evaluating the “reasoning quality” dimension.<\/p><\/blockquote>\n Whether you are using the new model graders or deterministic ones, the following best practices will help you get the most out of RFT.<\/p>\n RFT improves reasoning accuracy and decision quality in tasks where outputs can be clearly evaluated and scored. It is especially effective when:<\/p>\n Not a fit for style or tone.<\/strong> If you need formatting, voice, or stylistic adjustments, prefer prompt engineering, structured outputs, or supervised fine-tuning (SFT).<\/p><\/blockquote>\n Start by clearly stating the task and what success looks like. Then design a grader that reflects real task quality as reliably as possible. The grader is the primary driver of RFT success\u2014invest disproportionate effort here.<\/p>\n Before training, run a baseline evaluation on a small set of examples (10\u2013100 samples) so you understand starting performance and can measure real improvement. Evaluate using a base model (for example, o4-mini) and experiment with system prompts to reach the best possible performance before fine-tuning.<\/p>\n The grader determines what the model optimizes for. Follow these principles:<\/p>\n Begin with small datasets (10\u2013100 samples), simple graders, and low epoch counts. A practical workflow:<\/p>\n Expect RFT requires a different data format from SFT. The final message in each row must be a User<\/strong> or Developer<\/strong> role\u2014not Assistant.<\/p>\n SFT format<\/strong> (answer in the assistant message):<\/p>\n RFT format<\/strong> (answer moved to a top-level key for the grader):<\/p>\n The Every key referenced in your grader (e.g., Example of a mismatched grader:<\/strong><\/p>\n If your data uses To reference Treat tools as part of the environment, not passive helpers. Build tools that reflect the full decision-making cycle<\/strong> your task requires\u2014not just the final action. For example, an automatic escalation workflow shouldn’t only have a tool to trigger escalation; it also needs a tool to check recipient availability first. Without that step, the model never learns when<\/em> escalation is appropriate.<\/p>\n Design for training-scale traffic: set timeouts and rate limits, add tracing (latency + error codes), and plan retry behavior so that slow calls don’t cascade into a retry storm.<\/p>\n RFT supports tool use through function-calling, but MCP is the preferred approach for production agentic systems. Implement each tool once, then expose it two ways\u2014via an MCP interface for MCP-native clients and via a function-calling-compatible interface for fine-tuning. This lets you seamlessly integrate with Agents, Evaluations, and Reinforcement Fine-Tuning on the Foundry platform.<\/p>\n\n
![\"o4-mini](\"https:\/\/devblogs.microsoft.com\/foundry\/wp-content\/uploads\/sites\/89\/2026\/04\/ft-o4mini-global-training.webp\")
<\/p>\nCreate an o4-mini Global Training Job via REST API<\/h3>\n
curl -X POST \"https:\/\/<your-resource>.openai.azure.com\/openai\/fine_tuning\/jobs?api-version=2025-04-01-preview\" \\\r\n -H \"Content-Type: application\/json\" \\\r\n -H \"api-key: $AZURE_OPENAI_API_KEY\" \\\r\n -d '{\r\n \"model\": \"o4-mini\",\r\n \"training_file\": \"<your-training-file-id>\",\r\n \"method\": {\r\n \"type\": \"reinforcement\",\r\n \"reinforcement\": {\r\n \"grader\": {\r\n \"type\": \"string_check\",\r\n \"name\": \"answer-check\",\r\n \"input\": \"{{sample.output_text}}\",\r\n \"reference\": \"{{item.reference_answer}}\",\r\n \"operation\": \"eq\"\r\n }\r\n }\r\n },\r\n \"hyperparameters\": {\r\n \"n_epochs\": 2,\r\n \"compute_multiplier\": 1.0\r\n },\r\n \"trainingType\": \"globalstandard\"\r\n }'<\/code><\/pre>\n
\nNew Model Graders: GPT-4.1, GPT-4.1-mini, and GPT-4.1-nano<\/h2>\n
\n
When to Use Model Graders<\/h3>\n
\n
Choosing the Right Model Grader<\/h3>\n
\n
\nReinforcement Fine-Tuning Best Practices<\/h2>\n
When to Use RFT<\/h3>\n
\n
Step 1: Define the Objective<\/h3>\n
Step 2: Establish a Baseline<\/h3>\n
Step 3: Design Effective Graders<\/h3>\n
\n
Step 4: Start Small and Iterate<\/h3>\n
\n
Step 5: Tune Training Parameters<\/h3>\n
epoch count<\/code> and compute_multiplier<\/code> to have the most impact on quality. Adjust one at a time and monitor the reward trend and variance throughout training.<\/p>\n
\nRFT Data Format<\/h2>\n
{\r\n \"messages\": [\r\n { \"role\": \"system\", \"content\": \"Reply to the user's question as accurately as possible.\" },\r\n { \"role\": \"user\", \"content\": \"Question: What is the capital of France?\" },\r\n { \"role\": \"assistant\", \"content\": \"Paris\" }\r\n ]\r\n}<\/code><\/pre>\n{\r\n \"messages\": [\r\n { \"role\": \"developer\", \"content\": \"Reply to the user's question as accurately as possible.\" },\r\n { \"role\": \"user\", \"content\": \"Question: What is the capital of France?\" }\r\n ],\r\n \"reference_answer\": \"Paris\"\r\n}<\/code><\/pre>\nreference_answer<\/code> (or any custom top-level key) can be referenced in the grader as item.reference_answer<\/code>.<\/p>\n
\nCommon Pitfalls<\/h2>\n
Data and Grader Mismatch<\/h3>\n
item.reference_answer<\/code>) must exist in all<\/strong> data rows. If your grader references item.capital<\/code> but your data uses reference_answer<\/code>, the job will fail silently or score incorrectly.<\/p>\n{\r\n \"type\": \"string_check\",\r\n \"name\": \"answer-check\",\r\n \"input\": \"{{sample.output_text}}\",\r\n \"reference\": \"{{item.capital}}\",\r\n \"operation\": \"eq\"\r\n}<\/code><\/pre>\nreference_answer<\/code> instead of capital<\/code>, update the grader reference to {{item.reference_answer}}<\/code>.<\/p>\nMissing Response Format<\/h3>\n
sample.output_json<\/code> in your grader, you must provide a response format in the job definition. Without it, the model outputs free-form text and JSON grader references will fail.<\/p>\n{\r\n \"type\": \"json_schema\",\r\n \"json_schema\": {\r\n \"name\": \"response\",\r\n \"strict\": true,\r\n \"schema\": {\r\n \"properties\": {\r\n \"capital\": { \"title\": \"Capital\", \"type\": \"string\" },\r\n \"population\": { \"title\": \"Population\", \"type\": \"string\" }\r\n },\r\n \"title\": \"CapitalData\",\r\n \"type\": \"object\",\r\n \"additionalProperties\": false\r\n }\r\n }\r\n}<\/code><\/pre>\n
\nAdvanced: Agentic RFT Scenarios<\/h2>\n
Tool Design<\/h3>\n
MCP Server Integration<\/h3>\n