# Training and evaluation record

## v1 provenance

- Model: YOLO26x-Pose
- Version: `v1`
- Status: `pretrained` (inventory status); model card status remains `experimental` until independent validation completes
- Upstream trainer/developer: Ultralytics
- Upstream model reference: [Ultralytics YOLO26x-Pose](https://platform.ultralytics.com/ultralytics/yolo26/yolo26x-pose)
- Select AI training run: none
- Select AI fine-tuning run: none
- Selected checkpoint: YOLO26x-Pose upstream pretrained checkpoint
- Checkpoint path: `models/yolo26x-pose.pt`
- Provenance script: `scripts/train.py` (records provenance only; does not launch training or download datasets)

The v1 package contains the upstream pretrained YOLO26x-Pose artifact from Ultralytics. Select AI did not train or fine-tune the published checkpoint. Model specifications for the upstream checkpoint are taken from the Ultralytics platform reference linked above.

## Output contract (v1)

- Per-person feature vector: 51 floats (17 COCO keypoints × `x`, `y`, `confidence`)
- Per-person bounding box: `box_xyxy = [x1, y1, x2, y2]`
- Downstream consumers assemble bbox + feature vector into a pandas DataFrame

## Reference benchmark record

The retained benchmark record documents the upstream reference performance for YOLO26x-Pose at 960 × 960 on the COCO Keypoints validation set.

| Model | Input | Dataset | mAP50-95 | mAP50 |
| --- | --- | --- | ---: | ---: |
| YOLO26x-Pose | 960 × 960 | COCO Keypoints | 71.6% | 91.6% |