> ## 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 Training

> Configure and launch custom YOLOv11 training jobs.

# 🎛️ Configuring a Training Job

This guide explains every field on the **Train Model** panel to help you understand and configure your model training setup.

<Frame>
  <img src="https://mintcdn.com/ocularai/FeaFuA3bMytYuXie/images/model-training/train-model-panel-light.png?fit=max&auto=format&n=FeaFuA3bMytYuXie&q=85&s=94a2dfd2c3e148e4f8f3d19c7a5233fb" alt="Training Config UI" width="724" height="1590" data-path="images/model-training/train-model-panel-light.png" />
</Frame>

## 1. Details

| Field           | Purpose                         | Example                     |
| --------------- | ------------------------------- | --------------------------- |
| **Name**        | Identifies the run in history   | `Hand_Detector_v1_10Ep`     |
| **Description** | (optional) Notes for future you | `Testing larger image size` |

## 2. Base Model

You can currently choose between two base models depending on the project type:

1. **YOLOv11 Object Detection** – used with bounding boxes annotations
2. **YOLOv11 Segmentation** – used with polygon and SAM-2 generated annotations

Depending on the current project's type (object detection or segmentation), you will only see the one available model for that project type.

Both models the latest generation from Ultralytics, fusing an upgraded backbone
and neck for higher accuracy with fewer parameters \[Ultralytics Docs].

## 3. Model Variant

Choosing a variant is a trade-off between speed & accuracy.

| Variant | Params (M) | COCO mAP50-95² | Notes                      |
| ------- | ---------- | -------------- | -------------------------- |
| `n`     | 2.6        | 39.5           | Ultra-light, mobile & IoT  |
| `s`     | 9.4        | 47.0           | Good for edge GPUs         |
| `m`     | 20.1       | 51.5           | Balanced; default          |
| `l`     | 26.2       | 54.0           | Higher accuracy, more VRAM |
| `x`     | 58.8       | 56.8           | Maximum accuracy, slowest  |

<small>¹ Tesla T4 @ 640 px ² COCO val set</small>

<Check>`m` is an excellent starting point for most users.</Check>

## 4. Customization

| Option                           | When to use                          |
| -------------------------------- | ------------------------------------ |
| **From Scratch**                 | Fresh dataset or new architecture.   |
| **Upload Weights** (coming soon) | Resume / fine-tune a previous model. |

## 5. Hyper-parameters

| Parameter  | UI Control                | Range         | What it does                                                                          |
| ---------- | ------------------------- | ------------- | ------------------------------------------------------------------------------------- |
| Epochs     | Slider                    | 1-100         | Number of full passes through the training set.                                       |
| Image Size | Radio-buttons (320 / 640) | 320 or 640 px | Resolution the model is trained with. Higher helps small objects but needs more VRAM. |

<Note>
  Other YOLO hyper-parameters (batch size, learning rate, momentum, etc.) are
  managed automatically by the trainer and use Ultralytics' recommended
  defaults.
</Note>

## 6. Launch the Job

Hit **Create Model**. Your run queues and begins on our GPU fleet. You will be redirected to the **models** tab, where performance
metrics and graphs will be available after the training run is complete.

***

<Note>
  Prefer code? Head to [Notebooks](/build-ai/notebooks) for SDK-based examples.
</Note>