# Model Training Comparison

This repository includes the best available smoke and fire detection weights and the evaluation artifacts needed to inspect the completed run. Trained weights are committed as regular files (only `best.pt`, no checkpoints or pretrained base models).

## Available runs

| Model | Run | Best epoch | Precision | Recall | mAP50 | mAP50-95 | Best weight |
| --- | --- | ---: | ---: | ---: | ---: | ---: | --- |
| YOLO11s | `smoke_fire_yolo11s_v1-4` | 84 | 0.77349 | 0.72164 | 0.77499 | 0.45688 | `runs/detect/smoke_fire_yolo11s_v1-4/weights/best.pt` |
| YOLO26n | `smoke_fire_yolo26n_v2-3` | Not available | Not available | Not available | Not available | Not available | `runs/detect/smoke_fire_yolo26n_v2-3/weights/best.pt` |

The YOLO11s row uses epoch 84 from `results.csv`, selected by the highest `metrics/mAP50-95(B)`. Training completed 124 epochs before early stopping. The best observed mAP50 was 0.77720 at epoch 90, while the selected checkpoint maximizes mAP50-95.

The YOLO26n directory currently contains only `best.pt`; its `args.yaml` and `results.csv` are unavailable. Its detection quality therefore cannot be compared fairly with YOLO11s yet. The weight is included as a deployment-size baseline only.

## Included evidence

The YOLO11s run includes:

- Full epoch metrics in `results.csv`
- Training and validation summary in `results.png`
- Precision, recall, F1, and PR curves
- Raw and normalized confusion matrices
- Validation label and prediction previews
- Effective training parameters in `args.yaml`
- Best-checkpoint metadata in `best_point.json`

Periodic `epoch*.pt` checkpoints, `last.pt`, training batch previews, runtime server logs, and manual audit preview images are intentionally excluded. They are either reproducible intermediates or unrelated to comparing the best models.

## Reproducing evaluation

Run validation against the configured smoke/fire dataset:

```powershell
uv run fire-yolo val --weights runs/detect/smoke_fire_yolo11s_v1-4/weights/best.pt
```

To complete the YOLO26n comparison, rerun validation with its best weight and retain the generated `args.yaml`, `results.csv`, plots, and confusion matrices beside the checkpoint.