---
pretty_name: VG100K4CL
tags:
  - image-text-retrieval
  - contrastive-learning
  - webdataset
  - visual-genome
  - computer-vision
language:
  - en
size_categories:
  - 100K<n<1M
---

# VG100K4CL

`VG100K4CL` is the full fine-tuning dataset used in the LPCV 2026 Track 1 image-to-text retrieval pipeline.

It includes the training JSONL files used for contrastive fine-tuning.

The full code for dataset construction, unpacking is available here:

`https://github.com/jn12-29/LPCV-Track1-EfficientAI`

That repository contains the detailed dataset-building pipeline and the code used to train and evaluate models with this dataset.

## What this repository contains

This repository contains the full fine-tuning dataset used by this project, including:

- contrastive training JSONL files
- the data needed to reproduce the training setup used in the repository

## Intended use

Use this dataset if you want to:

- fine-tune `MobileCLIP2-B` or related retrieval models on the same training data
- reproduce the project training setup
- inspect or reuse the provided contrastive JSONL annotations
- rebuild or extend the dataset pipeline using the linked project code

You likely do not need this dataset if you only want to run inference with the exported ONNX model.

## Download

```bash
hf download jn12/VG100K4CL \
  --repo-type dataset \
  --local-dir ./data/VG100K4CL
```


## Data format

This repository includes the complete fine-tuning data used by the project.

At a high level, it contains:

- image data
- contrastive JSONL annotations for training

The training JSONL format is:

```json
{
  "image_path": "build_datasets/data/VG_100K/107914.jpg",
  "positives": ["red bus on the street", "bus parked near buildings"],
  "hard_negatives": ["blue bus on the street", "bus parked in the water"]
}
```

The linked project repository contains the full code path for:

- unpacking image shards
- building or regenerating dataset variants
- training with the provided JSONL files
- exporting and evaluating models trained on this dataset

## Notes

- `VG100K4CL` should be understood as a publishable training dataset, not only as an intermediate image-source dump.
- The project repository documents how the provided dataset variants were produced and which one was used for final training.
- This repository was prepared for the LPCV 2026 Track 1 retrieval workflow centered on `MobileCLIP2-B`.


## Citation

If you use this dataset, please cite the repository and link back to the project code:

Authors:

`Hui Xie, Jinyang Du, Jiacheng Wang, Xiaoze Ge, Fengjun Zhong, Yejun Zeng, Ruihao Gong#, Xiaoning Liu, Shenghao Jin, Jinyang Guo#, Xianglong Liu`

```bibtex
@misc{vg100k4cl2026,
  title        = {VG100K4CL},
  author       = {Hui Xie and Jinyang Du and Jiacheng Wang and Xiaoze Ge and Fengjun Zhong and Yejun Zeng and Ruihao Gong and Xiaoning Liu and Shenghao Jin and Jinyang Guo and Xianglong Liu},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/jn12/VG100K4CL}}
}
```