README.md

metadata

license: mit
task_categories:
  - video-text-retrieval
  - text-to-video
language:
  - en
tags:
  - video-retrieval
  - generative-retrieval
  - semantic-ids
  - text-to-video
size_categories:
  - 10K<n<100K

GRDR-TVR: Generative Recall, Dense Reranking for Text-to-Video Retrieval

This dataset contains the pre-extracted video features and trained model checkpoints for the GRDR (Generative Recall, Dense Reranking) framework for efficient Text-to-Video Retrieval (TVR).

📄 Paper

Generative Recall, Dense Reranking: Learning Multi-View Semantic IDs for Efficient Text-to-Video Retrieval

Conference: SIGIR 2026

Paper PDF | Code Repository

📊 Dataset Overview

This dataset includes three main components:

1. InternVideo2 Features (~3.4GB)

Pre-extracted video features using InternVideo2 encoder for four benchmark datasets:

• MSR-VTT: 10,000 videos (932MB)
• ActivityNet: 20,000 videos (1.1GB)
• DiDeMo: 10,464 videos (916MB)
• LSMDC: 1,000 movies, 118,081 clips (424MB)

Feature Details:

• Dimension: 512-d embeddings
• Format: Pickle files (.pkl) with {video_id: embedding} mappings
• Extraction: InternVideo2 (InternVL-2B) with temporal pooling

2. GRDR Model Checkpoints (~2GB)

Trained GRDR models (T5-small based) for all four datasets:

• MSR-VTT: 494MB
• ActivityNet: 498MB
• DiDeMo: 504MB
• LSMDC: 478MB

Checkpoint Components:

• best_model.pt - Complete model checkpoint
• best_model.pt.model - T5 encoder-decoder weights
• best_model.pt.videorqvae - Video RQ-VAE quantizer
• best_model.pt.code - Pre-computed semantic IDs
• best_model.pt.centroids - Codebook centroids
• best_model.pt.embedding - Learned embeddings
• best_model.pt.start_token - Start token embeddings

Model Architecture:

• Base: T5-small (60M parameters)
• Codebook size: 128/96/200 (dataset-dependent)
• Max code length: 3
• Training: 3-phase progressive training

3. Xpool Reranker Checkpoints (~7.2GB)

Pre-trained reranker models for dense reranking stage:

• MSR-VTT: msrvtt9k_model_best.pth (1.8GB)
• ActivityNet: actnet_model_best.pth (1.8GB)
• DiDeMo: didemo_model_best.pth (1.8GB)
• LSMDC: lsmdc_model_best.pth (1.8GB)

Reranker Details:

• Architecture: CLIP-based (ViT-B/32)
• Purpose: Fine-grained reranking of recalled candidates
• Format: PyTorch checkpoint files (.pth)

🚀 Quick Start

Download Specific Components

Using Python Script

# Download everything
python download_features.py --all

# Download only InternVideo2 features for specific datasets
python download_features.py --features --datasets msrvtt actnet

# Download GRDR checkpoints only
python download_features.py --grdr

# Download Xpool reranker only
python download_features.py --xpool --datasets msrvtt

Using Hugging Face CLI

# Download entire dataset
huggingface-cli download JasonCoderMaker/GRDR-TVR --repo-type dataset --local-dir ./GRDR-TVR

# Download specific component
huggingface-cli download JasonCoderMaker/GRDR-TVR InternVideo2/msrvtt --repo-type dataset --local-dir ./features

# Download GRDR checkpoint for MSR-VTT
huggingface-cli download JasonCoderMaker/GRDR-TVR GRDR/msrvtt --repo-type dataset --local-dir ./checkpoints

Load Features in Python

import pickle
from huggingface_hub import hf_hub_download

# Download and load InternVideo2 features
feature_file = hf_hub_download(
    repo_id="JasonCoderMaker/GRDR-TVR",
    filename="InternVideo2/msrvtt/msrvtt_internvideo2.pkl",
    repo_type="dataset"
)

with open(feature_file, 'rb') as f:
    video_features = pickle.load(f)
    
# Access features
video_id = "video7015"
embedding = video_features[video_id]  # Shape: (512,)
print(f"Feature shape: {embedding.shape}")

Load GRDR Model

import torch
from huggingface_hub import hf_hub_download

# Download checkpoint
checkpoint_path = hf_hub_download(
    repo_id="JasonCoderMaker/GRDR-TVR",
    filename="GRDR/msrvtt/best_model/best_model.pt",
    repo_type="dataset"
)

# Load model
checkpoint = torch.load(checkpoint_path, map_location='cpu')
print(f"Available keys: {checkpoint.keys()}")

# Use with your GRDR model
from models.grdr import GRDR
model = GRDR(...)
model.load_state_dict(checkpoint['model'], strict=False)

Load Xpool Reranker

import torch
from huggingface_hub import hf_hub_download

# Download reranker checkpoint
reranker_path = hf_hub_download(
    repo_id="JasonCoderMaker/GRDR-TVR",
    filename="Xpool/msrvtt9k_model_best.pth",
    repo_type="dataset"
)

# Load reranker
checkpoint = torch.load(reranker_path, map_location='cpu')
# Use with your Xpool model

📂 Repository Structure

GRDR-TVR/
├── README.md                          # This file
├── download_features.py               # Python download utility
├── download_checkpoints.sh            # Bash download script
│
├── InternVideo2/                      # Video Features (3.4GB)
│   ├── actnet/
│   │   └── actnet_internvideo2.pkl
│   ├── didemo/
│   │   └── didemo_internvideo2.pkl
│   ├── lsmdc/
│   │   └── lsmdc_internvideo2.pkl
│   └── msrvtt/
│       └── msrvtt_internvideo2.pkl
│
├── GRDR/                              # GRDR Checkpoints (2GB)
│   ├── actnet/best_model/
│   ├── didemo/best_model/
│   ├── lsmdc/best_model/
│   └── msrvtt/best_model/
│
└── Xpool/                             # Reranker Checkpoints (7.2GB)
    ├── actnet_model_best.pth
    ├── didemo_model_best.pth
    ├── lsmdc_model_best.pth
    └── msrvtt9k_model_best.pth

🔬 Dataset Statistics

Dataset	Videos	Train Queries	Test Queries	Feature Size	GRDR Size	Xpool Size
MSR-VTT	10,000	9,000	1,000	932 MB	494 MB	1.8 GB
ActivityNet	20,000	10,009	4,917	1.1 GB	498 MB	1.8 GB
DiDeMo	10,464	8,395	1,065	916 MB	504 MB	1.8 GB
LSMDC	118,081	118,081	1,000	424 MB	478 MB	1.8 GB
Total	-	-	-	3.4 GB	2.0 GB	7.2 GB

🎯 Performance

GRDR achieves competitive accuracy with dense retrievers while being significantly more efficient:

Dataset	R@1	R@5	R@10	Storage Reduction	Speed-up
MSR-VTT	45.2	72.1	81.3	15.6×	287×
ActivityNet	41.8	76.2	86.4	12.3×	310×
DiDeMo	43.1	71.8	81.7	18.2×	265×
LSMDC	24.3	48.9	59.2	22.1×	298×

Compared to CLIP4Clip baseline with exhaustive search

💻 Usage in GRDR Pipeline

Complete Retrieval Pipeline

from models.grdr import GRDR
from reranker.xpool import XpoolReranker
import torch

# 1. Load video features
video_features = load_internvideo2_features("msrvtt")

# 2. Load GRDR model for recall
grdr_model = GRDR.from_pretrained("JasonCoderMaker/GRDR-TVR", dataset="msrvtt")

# 3. Generate candidates (fast generative recall)
query = "A person playing guitar"
candidates = grdr_model.generate_candidates(query, top_k=100)

# 4. Load Xpool reranker
reranker = XpoolReranker.from_pretrained("JasonCoderMaker/GRDR-TVR", dataset="msrvtt")

# 5. Rerank candidates (dense reranking)
final_results = reranker.rerank(query, candidates, top_k=10)

🛠️ Requirements

pip install torch torchvision
pip install transformers>=4.30.0
pip install huggingface_hub
pip install sentencepiece

For the full GRDR codebase, see: GitHub Repository

📜 Citation

If you use this dataset or models in your research, please cite:

@inproceedings{grdr2026,
  title={Generative Recall, Dense Reranking: Learning Multi-View Semantic IDs for Efficient Text-to-Video Retrieval},
  author={Anonymous},
  booktitle={Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2026}
}

📝 License

This dataset is released under the MIT License. See LICENSE for details.

The video datasets (MSR-VTT, ActivityNet, DiDeMo, LSMDC) are subject to their original licenses. This repository only provides pre-extracted features, not the original videos.

🙏 Acknowledgments

• InternVideo2: We thank the authors of InternVideo2 for their excellent video encoder
• Xpool: The reranker architecture is based on X-POOL
• Datasets: MSR-VTT, ActivityNet Captions, DiDeMo, and LSMDC benchmark creators

📧 Contact

For questions or issues, please open an issue on the GitHub repository or contact the authors.

---

Dataset Version: 1.0
Last Updated: January 2026
Maintained by: @JasonCoderMaker