View Training Results

About

Roboflow produces training graphs and summary metrics for every model you train on the platform, letting you monitor a job while it runs and evaluate the resulting model once it finishes. The metrics shown depend on your project type - precision, recall, and mAP for object detection; accuracy for classification; mAP for segmentation and keypoint models; and perplexity for multimodal models. You can also test a finished model on sample or custom images before wiring it into an application.

Training graphs let you see the status of your model during training. Training graphs are available for all models trained on Roboflow.

You will also see metrics that summarise your model performance.

The metrics you will see will depend on your model type:

• Object detection projects show the precision, recall, and mAP of the model.
• Classification projects show accuracy.
• Segmentation and keypoint models show mAP score.
• Multimodal models show perplexity.

During Training

After you start training a model, a message will appear on the dataset version page associated with the model you are training. This message will first say that a training machine is starting.

If a version has more than one training, select a training on the version page to view its graphs and metrics.

You will then see training graphs appear live as your model trains:

NAS Training Charts

For models trained with Neural Architecture Search (NAS) on object detection or instance segmentation projects, a specialized paired chart view is displayed instead of standard training graphs:

• Training Progress Band: A gradient-filled band showing the range of model performance (upper and lower bounds) over training epochs. Hover over any epoch to inspect metrics at that point. A pulsing indicator shows the current training front during live training.
• Epoch Snapshot: A Pareto curve showing the trade-off between model latency and accuracy for the architectures evaluated at each epoch. The gradient spectrum ranges from faster (lighter) to more accurate (darker) models.

Unlike Standard and Plus NAS runs, Fast NAS runs have no baseline configurations to report a spread of metrics for during training, so they report metrics only for the single supernet being trained. These runs show the standard training graph instead of the paired band and snapshot view.

After Training

When your model has finished training, you can view metrics from the training process.

To find your training graphs for a model after training, first click the Models tab in the sidebar:

Then, click on the training (or model) whose training graphs you want to see.

Scroll down on the page until you see the Training Graphs section:

Test Your Model

You can test your model on an image from the Test page. This is a good way to run quick checks to see how your model performs before you build your application logic in Workflows.

The Test feature works for object detection, segmentation, classification, and keypoint models. Test is not supported for multimodal projects.

To test a model, click "Test" on the right sidebar. The Test page will then open:

The Test page will show several images from your test set from which you can select to run on your model. You can also upload your own images and videos, or try with your webcam.

To find deployment instructions for your model, click Try on My Machine. You can also deploy your model on Roboflow Workflows.

HTTP API

Get Training Results

Retrieve the training results for a dataset version. For NAS (Neural Architecture Search) training runs, this returns a bundled view including all child models, their metrics, hardware-specific recommendations, and mining frontier data. For standard fine-tune runs, it returns a minimal summary.

This endpoint consolidates multiple reads into a single call, so you don't need to fetch model data separately.

Endpoint

GET /:workspace/:project/:version/training/results

Authentication: Requires an API key with version.read scope.

Request

curl "https://api.roboflow.com/{workspace}/{project}/{version}/training/results?api_key=$ROBOFLOW_API_KEY"

Response (NAS Training)

{
  "trainingId": "my-workspace/my-project/5",
  "versionId": "5",
  "status": "finished",
  "jobType": "nas",
  "modelType": "rfdetr-nas-parent",
  "modelGroup": "rfdetrNasGroup-2026-04-01",
  "modelCount": 8,
  "recommendedByHardware": {
    "jetson-orin-nano": "nas-child-1",
    "t4": "nas-child-2"
  },
  "mining": {
    "mining": {
      "progress": 0.42,
      "frontier": [{"name": "subnet-1", "mAP": 0.81}]
    },
    "baseline": [{"modelTypeDisplay": "rfdetr-nas-S"}]
  },
  "models": [
    {
      "modelId": "nas-child-1",
      "modelType": "rfdetr-nas-S",
      "status": "finished",
      "nasFamily": "child",
      "metrics": {
        "map50": 81,
        "map5095": 64,
        "f1": 78,
        "hardware": "jetson-orin-nano",
        "latency": 12.4,
        "paretoOptimalFor": ["jetson-orin-nano"]
      },
      "recommended": true
    }
  ]
}

Response (Fine-Tune Training)

For non-NAS training runs, the response is a minimal bundle:

{
  "trainingId": "my-workspace/my-project/3",
  "versionId": "3",
  "status": "finished",
  "jobType": "fine-tune",
  "modelGroup": null,
  "modelCount": 0,
  "models": []
}

Response Fields

Model summary fields:

Errors

MCP Server

Connect your AI agent to the MCP Server and it can check how training went with these tools: