---
license: etalab-2.0
---

## Model Weights

The trained weights for all benchmarks are hosted on [Hugging Face](https://huggingface.co/datasets/anonymous-submission-dataset-code/TiBuDB).

### Weights Organization

Download the weights and place them in the `TiBuDB_trained_weights/` directory.

| Task | Model | Weight File | Description | SAHI Crop Size | Inference Size |
| :--- | :--- | :--- | :--- | :--- | :--- |
| **Detection** | YOLO26x | `best_det_yolo26x_seed1000_baseline.pt` | Baseline (1x) | 128 | 128 |
| **Detection** | YOLO26x | `best_det_yolo26x_seed1000_x4.pt` | Upscaled (4x) | 128 | 512 |
| **Detection** | RF-DETR | `best_ema_det_rfdetr_large_seed0_baseline.pth` | Transformer Baseline | 128 | N/A |
| **Segmentation** | YOLO26x | `best_seg_yolo26x_seed100_baseline.pt` | Baseline (1x) | 128 | 128 |
| **Segmentation** | YOLO26x | `best_seg_yolo26x_seed100_x4.pt` | Upscaled (4x) | 128 | 512 |
| **Segmentation** | RF-DETR | `best_ema_seg_rfdetr_large_seed100_baseline.pth` | Transformer Baseline | 128 | N/A |
| **OBB** | YOLO26x | `best_obb_yolo26x_seed5000_baseline.pt` | Oriented Bbox (1x) | 128 | 128 |
| **OBB** | YOLO26x | `best_obb_yolo26x_seed5000_x4.pt` | Oriented Bbox (4x) | 128 | 512 |

> **Note:** RF-DETR processes images at the native crop size (128) without upscaling; inference size is not applicable.

### Quick Load Example

**Ultralytics (YOLO / RT-DETR)**

```python
from ultralytics import YOLO

model = YOLO("TiBuDB_trained_weights/best_det_yolo26x_seed1000_baseline.pt")
results = model.predict("path/to/image.png")
```

**RF-DETR**

```python
from rfdetr import RFDETRLarge

model = RFDETRLarge(pretrain_weights="TiBuDB_trained_weights/best_ema_det_rfdetr_large_seed0_baseline.pth")
results = model.predict("path/to/image.png")
```