README.md

---
license: cc-by-4.0
pretty_name: LightCMR-Bench
task_categories:
- image-feature-extraction
- text-retrieval
language:
- en
size_categories:
- 100K<n<1M
tags:
- cross-modal-retrieval
- image-text-retrieval
- product-quantization
- faiss
- beit3
- localized-narratives
viewer: false
---

# LightCMR-Bench

LightCMR-Bench is a feature-level cross-modal retrieval benchmark and reviewer artifact package for TGRC-PQ. It provides pre-extracted image and text embeddings derived from the Localized Narratives benchmark, together with supplementary files used to reproduce the TGRC-PQ experiments.

This repository is not a `load_dataset()`-style tabular dataset. The Hugging Face Dataset Viewer is intentionally disabled because the repository contains NumPy feature matrices, Faiss indexes, PyTorch checkpoints, JSON/CSV result files, and qualitative evidence files.

## What Is Included

```text
embeddings_large_d1024/
  {coco_retrieval,flickr30k,open}/
    {train,val,test}/
      image.npy        # 1024D image features
      text.npy         # 1024D text features
      ids.json         # sample identifiers

embeddings_mini_d64/
  {coco_retrieval,flickr30k,open}/
    {train,val,test}/
      image.npy        # 64D image features
      text.npy         # 64D text features
      ids.json         # sample identifiers

openreview_artifacts/
  README.md
  faiss_models/                         # saved PQ/OPQ/RQ Faiss indexes and metadata
  checkpoints/locked_main_table/         # TGRC-PQ checkpoints and per-run logs
  supplementary_experiments_raw/         # raw ablation and sensitivity outputs
  qualitative_rank_evidence/             # ranking evidence and selected examples
  results/                               # compact summaries used by the paper
```

The previous root-level `coco_retrieval/`, `flickr30k/`, and `open/` folders have been removed. Please use the namespaced embedding folders above.

## Data Provenance

The files in this repository are derived feature files, not the original images or captions. The underlying image-text pairs come from the Localized Narratives benchmark:

<https://google.github.io/localized-narratives/>

We extracted 64D student features and 1024D teacher features with BEiT3-Large from the Flickr30K, COCO, and OpenImages portions of Localized Narratives, and use these feature files to build LightCMR-Bench for cross-modal retrieval and quantized-retrieval experiments.

Please cite the original Localized Narratives paper when using these artifacts. This repository should not be treated as a replacement for the original dataset or its license terms.

## Quick Download

Download only the 64D features:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="zhoukun/LightCMR-Bench",
    repo_type="dataset",
    allow_patterns="embeddings_mini_d64/**",
    local_dir="LightCMR-Bench",
)
```

Download only the 1024D features:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="zhoukun/LightCMR-Bench",
    repo_type="dataset",
    allow_patterns="embeddings_large_d1024/**",
    local_dir="LightCMR-Bench",
)
```

Download the TGRC-PQ reviewer artifacts:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="zhoukun/LightCMR-Bench",
    repo_type="dataset",
    allow_patterns="openreview_artifacts/**",
    local_dir="LightCMR-Bench",
)
```

## Using the Artifact Bundle with the Code Repository

In the TGRC-PQ code repository, create symlinks to the downloaded artifact folders:

```bash
mkdir -p data
ln -s /path/to/LightCMR-Bench data/lightcmr_bench
ln -s /path/to/LightCMR-Bench/openreview_artifacts/faiss_models faiss_models
```

For checkpoint inspection:

```bash
mkdir -p runs/locked_main_table
ln -s /path/to/LightCMR-Bench/openreview_artifacts/checkpoints/locked_main_table runs/locked_main_table/calibrated
```

The feature cache under `runs/faiss_feature_cache` is intentionally not included. It is deterministic cache data produced by the training script and can be regenerated from the feature files and Faiss models.

## File Sizes

Approximate remote sizes:

```text
embeddings_large_d1024: 6.39 GiB
embeddings_mini_d64:   0.58 GiB
openreview_artifacts:  0.16 GiB
```

## Citation

TGRC-PQ paper:

```text
Zhou, Kun and HASSAN, FADRATUL HAFINAZ and Gan, Keng Hoon,
Attention-Guided Product Quantization for Efficient Cross-Modal Retrieval.
Available at SSRN: https://ssrn.com/abstract=6022619 or
http://dx.doi.org/10.2139/ssrn.6022619
```

```bibtex
@misc{zhou_attention_guided_pq_ssrn,
  title        = {Attention-Guided Product Quantization for Efficient Cross-Modal Retrieval},
  author       = {Zhou, Kun and Hassan, Fadratul Hafinaz and Gan, Keng Hoon},
  howpublished = {Available at SSRN: \url{https://ssrn.com/abstract=6022619}},
  doi          = {10.2139/ssrn.6022619}
}
```

Localized Narratives:

```bibtex
@inproceedings{ponttuset2020connecting,
  title     = {Connecting Vision and Language with Localized Narratives},
  author    = {Pont-Tuset, Jordi and Uijlings, Jasper and Changpinyo, Soravit and Soricut, Radu and Ferrari, Vittorio},
  booktitle = {European Conference on Computer Vision},
  pages     = {647--664},
  year      = {2020},
  organization = {Springer}
}
```

LightCMR-Bench / TGRC-PQ artifacts:

```bibtex
@misc{zhoukun_lightcmr_bench,
  title        = {LightCMR-Bench: Feature-Level Cross-Modal Retrieval Benchmark and TGRC-PQ Artifacts},
  author       = {Zhoukun},
  howpublished = {\url{https://huggingface.co/datasets/zhoukun/LightCMR-Bench}},
  note         = {Derived features from Localized Narratives extracted with BEiT3-Large}
}
```