README.md

---
license: mit
task_categories:
- visual-document-retrieval
- text-to-video
language:
- en
tags:
- video-retrieval
- generative-retrieval
- semantic-ids
- text-to-video
size_categories:
- 1M<n<10M
---

# GRDR-TVR release assets

This dataset repository hosts release assets for GRDR: a learned semantic-key index for large-scale text-to-video retrieval. The main code repository is [JasonCoderMaker/GRDR](https://github.com/JasonCoderMaker/GRDR).

## Hosted files

### InternVideo2 features

Pre-extracted InternVideo2 features are available for:

- `msrvtt`
- `actnet`
- `didemo`
- `panda`
- `lsmdc` legacy assets

The source videos remain under their original dataset licenses and are not mirrored here.

### GRDR checkpoints

The release-facing checkpoint path is canonical for each dataset:

```text
GRDR/<dataset>/best_model/best_model.pt
GRDR/<dataset>/best_model/best_model.pt.code
GRDR/<dataset>/best_model/best_model.pt.centroids
GRDR/<dataset>/best_model/best_model.pt.embedding
GRDR/<dataset>/best_model/best_model.pt.model
GRDR/<dataset>/best_model/best_model.pt.start_token
GRDR/<dataset>/best_model/best_model.pt.videorqvae
```

Current release datasets:

- `msrvtt`
- `actnet`
- `didemo`
- `panda`

Legacy `lsmdc` checkpoint files remain available under `GRDR/lsmdc/best_model/`, but they are not part of the current GRDR release manifest.

### X-Pool checkpoints

X-Pool checkpoint filenames keep their original flat layout:

```text
Xpool/msrvtt9k_model_best.pth
Xpool/actnet_model_best.pth
Xpool/didemo_model_best.pth
Xpool/panda_2150k_s42_model_best.pth
Xpool/lsmdc_model_best.pth
```

## Download examples

Use the `download_features.py` script from the GRDR code repository for the full reproduction workflow. This Hub-side copy downloads the hosted release assets directly:

```bash
python download_features.py --features --datasets panda
python download_features.py --grdr --datasets msrvtt panda
python download_features.py --xpool --datasets msrvtt actnet didemo panda
python download_features.py --all --datasets panda
```

You can also download directly from the Hub:

```bash
hf download JasonCoderMaker/GRDR-TVR --type dataset --include 'InternVideo2/panda/**' --local-dir ./dataset/features
hf download JasonCoderMaker/GRDR-TVR --type dataset --include 'GRDR/panda/best_model/*' --local-dir ./output/checkpoints
hf download JasonCoderMaker/GRDR-TVR --type dataset --include 'Xpool/panda_2150k_s42_model_best.pth' --local-dir ./reranker/xpool/ckpt
```

The GRDR repository expects checkpoints under `output/checkpoints/GRDR/<dataset>/best_model/`.

## License

The release assets are distributed under the MIT License where applicable. Source video datasets keep their original licenses. This repository provides features and model checkpoints, not the original videos.