> ## Documentation Index
> Fetch the complete documentation index at: https://docs.useocular.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Model Evaluation

> Interpret metrics, graphs, and confusion matrices to judge model performance.

# 📈 Understanding Training Results

Once your training job completes, a rich set of metrics and visualizations will be available.
This page explains what they mean and how to act on them.

## 1. Performance Metrics

<Frame>
  <img src="https://mintcdn.com/ocularai/FeaFuA3bMytYuXie/images/model-training/model-metrics1.png?fit=max&auto=format&n=FeaFuA3bMytYuXie&q=85&s=cb222cc25b17024f73bdc8588b13e6dd" alt="Performance Metrics" width="1256" height="966" data-path="images/model-training/model-metrics1.png" />
</Frame>

| Metric        | Question it answers                       | Good value |
| ------------- | ----------------------------------------- | ---------- |
| **Precision** | How many predicted boxes are correct?     | >0.9       |
| **Recall**    | How many ground-truth objects were found? | >0.9       |
| **mAP50**     | Average precision @ IoU 0.50              | 0.85–0.95  |
| **mAP50-95**  | mAP averaged over 0.50–0.95 IoU           | 0.45–0.65  |

<Tip>
  `mAP50` is a high-level accuracy gauge, while `mAP50-95` evaluates box
  tightness.
</Tip>

### Metrics Across Epochs

Look for curves that rise then plateau—an indicator that learning has
stabilized. Sharp drops or spikes often mean an overly aggressive learning
rate.

## 2. Training & Validation Loss

Loss is the raw error signal the model is trying to minimise.

<Frame>
  <img src="https://mintcdn.com/ocularai/FeaFuA3bMytYuXie/images/model-training/loss-graphs.png?fit=max&auto=format&n=FeaFuA3bMytYuXie&q=85&s=d55e68d0dc3cc2889d156315e35fc2d7" alt="Loss across Epochs" width="972" height="1162" data-path="images/model-training/loss-graphs.png" />
</Frame>

* **Training Loss** should trend downwards steadily.
* **Validation Loss** should track the training curve. If it diverges upward
  your model is over-fitting.

## 3. Confusion Matrix

Rows = *actual* class, columns = *predicted* class. Diagonal cells are correct
predictions; off-diagonal cells highlight confusions.

<Frame>
  <img src="https://mintcdn.com/ocularai/FeaFuA3bMytYuXie/images/model-training/confusion-matrix.png?fit=max&auto=format&n=FeaFuA3bMytYuXie&q=85&s=dbefeb2fa33fb235342f1928567854cf" alt="Confusion Matrix" width="1856" height="766" data-path="images/model-training/confusion-matrix.png" />
</Frame>

<Check>
  Hover a cell to reveal exact counts and click to view example frames in
  context.
</Check>

***

# 🛠️ Troubleshooting Checklist

| Symptom                                     | Likely Cause            | Remedy                                   |
| ------------------------------------------- | ----------------------- | ---------------------------------------- |
| **Validation loss ↑ while training loss ↓** | Over-fitting            | Add data, early-stop                     |
| **High precision, low recall**              | Conservative confidence | Lower NMS-conf, larger model             |
| **Low precision, high recall**              | Noisy labels            | Clean data, raise confidence             |
| **Both low**                                | Model too small         | Try `l` or `xl` variant, increase epochs |

## FAQ

**Why is mAP50 high but mAP50-95 low?**\
Boxes are in the right area but not tight. Try using more epochs.

**Why are there many background → class errors?**\
Add more pure background frames to train with

**When should I stop training?**\
When `val/mAP50` flattens or `Early-stopping` triggers after `patience` epochs.

***

<Check>
  Happy with the metrics? Download or deploy your model. Otherwise tweak settings
  in the [Training page](/build-ai/model-training) and try again.
</Check>