Supervised fine-tuning (SFT) lets you train an OpenAI model with examples for your specific use case. The result is a customized model that more reliably produces your desired style and content.
OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months.
All fine-tuned models will remain available for inference until their base models are deprecated. The full timeline is here.
| How it works | Best for | Use with |
|---|---|---|
Provide examples of correct responses to prompts to guide the model’s behavior. Often uses human-generated “ground truth” responses to show the model how it should respond. |
|
|
Supervised fine-tuning has four major parts:
- Build your training dataset to determine what “good” looks like
- Upload a training dataset containing example prompts and desired model output
- Create a fine-tuning job for a base model using your training data
- Evaluate your results using the fine-tuned model
Good evals first! Only invest in fine-tuning after setting up evals. You need a reliable way to determine whether your fine-tuned model is performing better than a base model.
Build a robust, representative dataset to get useful results from a fine-tuned model. Use the following techniques and considerations.
Right number of examples
- The minimum number of examples you can provide for fine-tuning is 10
- We see improvements from fine-tuning on 50–100 examples, but the right number for you varies greatly and depends on the use case
- We recommend starting with 50 well-crafted demonstrations and evaluating the results
If performance improves with 50 good examples, try adding examples to see further results. If 50 examples have no impact, rethink your task or prompt before adding training data.
What makes a good example
- Whatever prompts and outputs you expect in your application, as realistic as possible
- Specific, clear questions and answers
- Use historical data, expert data, logged data, or other types of collected data
Formatting your data
- Use JSONL format, with one complete JSON structure on every line of the training data file
- Use the chat completions format
- Your file must have at least 10 lines
JSONL format example file
An example of JSONL training data, where the model calls a get_weather function:
{"messages":[{"role":"user","content":"What is the weather in San Francisco?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. San Francisco, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Minneapolis?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Minneapolis, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Minneapolis, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in San Diego?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Diego, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. San Diego, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Memphis?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Memphis, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Memphis, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Atlanta?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Atlanta, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Atlanta, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Sunnyvale?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Sunnyvale, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Sunnyvale, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Chicago?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Chicago, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Chicago, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Boston?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Boston, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Boston, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in Honolulu?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Honolulu, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. Honolulu, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}
{"messages":[{"role":"user","content":"What is the weather in San Antonio?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Antonio, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. San Antonio, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}Distilling from a larger model
One way to build a training data set for a smaller model is to distill the results of a large model to create training data for supervised fine tuning. The general flow of this technique is:
- Tune a prompt for a larger model (like
gpt-4.1) until you get great performance against your eval criteria. - Capture results generated from your model using whatever technique is convenient - note that the Responses API stores model responses for 30 days by default.
- Use the captured responses from the large model that fit your criteria to generate a dataset using the tools and techniques described above.
- Tune a smaller model (like
gpt-4.1-mini) using the dataset you created from the large model.
This technique can enable you to train a small model to perform similarly on a specific task to a larger, more costly model.
Upload your dataset of examples to OpenAI. We use it to update the model’s weights and produce outputs like the ones included in your data.
In addition to text completions, you can train the model to more effectively generate structured JSON output or function calls.
Upload your data with button clicks
- Navigate to the dashboard > fine-tuning.
- Click + Create.
- Under Training data, upload your JSONL file.
With your test data uploaded, create a fine-tuning job to customize a base model using the training data you provide. When creating a fine-tuning job, you must specify:
- A base model (
model) to use for fine-tuning. This can be either an OpenAI model ID or the ID of a previously fine-tuned model. See which models support fine-tuning in the model docs. - A training file (
training_file) ID. This is the file you uploaded in the previous step. - A fine-tuning method (
method). This specifies which fine-tuning method you want to use to customize the model. Supervised fine-tuning is the default.
Upload your data with button clicks
- In the same + Create modal as above, complete the required fields.
- Select supervised fine-tuning as the method and whichever model you want to train.
- When you’re ready, click Create to start the job.
Use the approaches below to check how your fine-tuned model performs. Adjust your prompts, data, and fine-tuning job as needed until you get the results you want. The best way to fine-tune is to continue iterating.
Compare to evals
To see if your fine-tuned model performs better than the original base model, use evals. Before running your fine-tuning job, carve out data from the same training dataset you collected in step 1. This holdout data acts as a control group when you use it for evals. Make sure the training and holdout data have roughly the same diversity of user input types and model responses.
Learn more about running evals.
Monitor the status
Check the status of a fine-tuning job in the dashboard or by polling the job ID in the API.
Monitor in the UI
- Navigate to the fine-tuning dashboard.
- Select the job you want to monitor.
- Review the status, checkpoints, message, and metrics.
Try using your fine-tuned model
Evaluate your newly optimized model by using it! When the fine-tuned model finishes training, use its ID in either the Responses or Chat Completions API, just as you would an OpenAI base model.
Use your model in the Playground
- Navigate to your fine-tuning job in the dashboard.
- In the right pane, navigate to Output model and copy the model ID. It should start with
ft:… - Open the Playground.
- In the Model dropdown menu, paste the model ID. Here, you should also see other fine-tuned models you’ve created.
- Run some prompts and see how your fine-tuned performs!
Use checkpoints if needed
Checkpoints are models you can use. We create a full model checkpoint for you at the end of each training epoch. They’re useful in cases where your fine-tuned model improves early on but then memorizes the dataset instead of learning generalizable knowledge—called _overfitting. Checkpoints provide versions of your customized model from various moments in the process.
Find checkpoints in the dashboard
- Navigate to the fine-tuning dashboard.
- In the left panel, select the job you want to investigate. Wait until it succeeds.
- In the right panel, scroll to the list of checkpoints.
- Hover over any checkpoint to see a link to launch in the Playground.
- Test the checkpoint model’s behavior by prompting it in the Playground.
Currently, only the checkpoints for the last three epochs of the job are saved and available for use.
Before launching in production, review and follow the following safety information.
Once a fine-tuning job is completed, we assess the resulting model’s behavior across 13 distinct safety categories. Each category represents a critical area where AI outputs could potentially cause harm if not properly controlled.
| Name | Description |
|---|---|
| advice | Advice or guidance that violates our policies. |
| harassment/threatening | Harassment content that also includes violence or serious harm towards any target. |
| hate | Content that expresses, incites, or promotes hate based on race, gender, ethnicity, religion, nationality, sexual orientation, disability status, or caste. Hateful content aimed at non-protected groups (e.g., chess players) is harassment. |
| hate/threatening | Hateful content that also includes violence or serious harm towards the targeted group based on race, gender, ethnicity, religion, nationality, sexual orientation, disability status, or caste. |
| highly-sensitive | Highly sensitive data that violates our policies. |
| illicit | Content that gives advice or instruction on how to commit illicit acts. A phrase like “how to shoplift” would fit this category. |
| propaganda | Praise or assistance for ideology that violates our policies. |
| self-harm/instructions | Content that encourages performing acts of self-harm, such as suicide, cutting, and eating disorders, or that gives instructions or advice on how to commit such acts. |
| self-harm/intent | Content where the speaker expresses that they are engaging or intend to engage in acts of self-harm, such as suicide, cutting, and eating disorders. |
| sensitive | Sensitive data that violates our policies. |
| sexual/minors | Sexual content that includes an individual who is under 18 years old. |
| sexual | Content meant to arouse sexual excitement, such as the description of sexual activity, or that promotes sexual services (excluding sex education and wellness). |
| violence | Content that depicts death, violence, or physical injury. |
Each category has a predefined pass threshold; if too many evaluated examples in a given category fail, OpenAI blocks the fine-tuned model from deployment. If your fine-tuned model does not pass the safety checks, OpenAI sends a message in the fine-tuning job explaining which categories don’t meet the required thresholds. You can view the results in the moderation checks section of the fine-tuning job.
How to pass safety checks
In addition to reviewing any failed safety checks in the fine-tuning job object, you can retrieve details about which categories failed by querying the fine-tuning API events endpoint. Look for events of type moderation_checks for details about category results and enforcement. This information can help you narrow down which categories to target for retraining and improvement. The model spec has rules and examples that can help identify areas for additional training data.
While these evaluations cover a broad range of safety categories, conduct your own evaluations of the fine-tuned model to ensure it’s appropriate for your use case.
Now that you know the basics of supervised fine-tuning, explore these other methods as well.
Learn to fine-tune for computer vision with image inputs.
Direct preference optimization
Fine-tune a model using direct preference optimization (DPO).
Reinforcement fine-tuning
Fine-tune a reasoning model by grading its outputs.