---
license: other
license_name: original-dataset-and-image-terms
license_link: https://huggingface.co/datasets/jeffry77/track_dataset#original-sources-and-rights
---
# TrackingAnything native datasets

Original-format corpus mirror, not model-generated annotations or converted dialogs.
The original **205-file training/shared-data release is complete**; its
`release_manifest.json`, `checksums.json`, archive paths and bytes are unchanged.

## Separate test extension

Status: **complete and remotely hash-verified**. The added TNL2K/VastTrack test images and native annotations
have a separate `test_release_manifest.json` and, after completion, `test_checksums.json`.
Original training manifests intentionally do not list these test-only additions.

| Dataset | Training inputs (unchanged) | Test inputs | Details |
|---|---|---|---|
| TNL2K | archives/train: 14 ZIPs; official annotation patches | archives/test: 9 independent TAR.GZ archives, 700 sequences | [TNL2K](TNL2K/README.md) |
| VastTrack | archives/train: 82 ZIPs, 47,109 supplied native sequences (paper: 47,110) | archives/test: 15 ZIPs, 3500 sequences | [VastTrack](VastTrack/README.md) |
| LaSOT | archives: 70 mixed-split category ZIPs | Same archives, official 1120/280 split | [LaSOT](LaSOT/README.md) |
| TNLLT | archives: 30 contiguous volumes | Same archives, official 150/50 split | [TNLLT](TNLLT/README.md) |
| RefCOCO/+/g | Shared train2014 images and UNC/UNC/UMD annotations | Annotation-defined val/test using shared images | [RefCOCO](RefCOCO/README.md) |

Archives include native annotations where supplied, not only pixels. Test membership
must not be merged into training. No Stage0 cache, private credentials or model weights
are distributed here. Adding test files does not change the existing training recipe.
The current `align_hf.py` still consumes only the original training manifest and its
declared training-release scope; test extraction/evaluation is separate.

```bash
# Original training inputs and shared train/test source packages:
hf download jeffry77/track_dataset --repo-type dataset --local-dir ./dataset --exclude "TNL2K/archives/test/*" "VastTrack/archives/test/*" "test_*"
# Test archives, test release metadata and per-dataset layout instructions:
hf download jeffry77/track_dataset --repo-type dataset --local-dir ./dataset --include "TNL2K/archives/test/*" "VastTrack/archives/test/*" "test_*" "*/README.md"
```

## Original sources and rights

Original dataset and underlying image licenses remain applicable; this mirror grants
no additional rights and implies no author endorsement. Preserve the original splits.

- [TNL2K](https://github.com/wangxiao5791509/TNL2K_evaluation_toolkit)
- [LaSOT](https://hengfan2010.github.io/projects/LaSOT/)
- [VastTrack](https://github.com/HengLan/VastTrack)
- [TNLLT](https://github.com/Event-AHU/Open_VLTrack)
- [REFER / RefCOCO](https://github.com/lichengunc/refer)

Integrity: completed payloads are checked by size and cloud SHA256/Git blob hashes.
Test archives also undergo full ZIP CRC or gzip/TAR streaming validation locally.
This does not independently certify every semantic annotation or pixel's truth.