---
license: cc-by-4.0
pretty_name: CS-CLIP Training Annotations
language:
- en
tags:
- image-text-retrieval
- compositionality
- clip
- arxiv:2602.23906
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# CS-CLIP Training Annotations

Prepared training annotations for **[Half-Truths Break Similarity-Based Retrieval](https://arxiv.org/abs/2602.23906)**, NeurIPS 2026. Bora Kargi, Arnas Uselis, Seong Joon Oh.

This release contains **410,340 caption records** from the eight annotation files used by the reported CS-CLIP run. All reference COCO `train2014` images. Caption counts are not unique image counts. Images are downloaded separately from COCO.

The [verified Half-Truth evaluation dataset](https://huggingface.co/datasets/kbora/Half-Truths) is released separately.

## Train using the original files

The archive preserves the training JSON files exactly, including their original field names. The [CS-CLIP repository](https://github.com/kargibora/CS-CLIP/tree/release/paper-reproduction) provides training code.

```bash
hf download kbora/CS-CLIP-Training original/training-json.tar.gz \
  --repo-type dataset --local-dir datasets/CS-CLIP-Training
mkdir -p datasets/CS-CLIP-Training/json
tar -xzf datasets/CS-CLIP-Training/original/training-json.tar.gz \
  -C datasets/CS-CLIP-Training/json
```

`image_path` is relative to the image root, for example `datasets/COCO/train2014/COCO_train2014_000000057870.jpg`. With this directory layout, set `IMAGE_ROOT=.`.

## Browse the annotations

```python
from datasets import load_dataset
samples = load_dataset("kbora/CS-CLIP-Training", split="train")
```

The Parquet view exposes the main training fields in a tabular schema:

- `sample_id`, `original_caption`, `image_path`: caption and image references.
- `entities`: extracted positive units, called `positive_components` in the original JSON.
- `entity_foils`: rows of `positive`, `negative`, and `change_type`, flattened from `negative_components`.
- `relations_json`: JSON-encoded relation units and their matched foils, preserving nested source fields.
- `swap_negatives`: full-caption shuffled negatives.

Use the original archive for exact training inputs; the Parquet files are a browsing view. `manifest.json` records each original file's SHA-256 and sample count.

## Construction and limitations

The annotations contain automatically generated entity/relation units, matched foils, and shuffled caption negatives. They are **not human-verified training labels**. An image-grounded VLM audit of 1,000 training foils estimated a 24.6% false-negative rate: some nominally incorrect foils are true for the image. The paper's evaluation uses a separate human-verified suite. The training archive is distributed as used, including that noise.

## Source assets and licensing

The authors’ generated annotations are released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). This grant covers their contributions, not third-party source captions or images.

Source captions and images come from [COCO](https://cocodataset.org/). COCO images are not included in this repository and retain their original rights and terms. An annotation license does not relicense source assets.

## Citation

```bibtex
@inproceedings{kargi2026halftruths,
  title={Half-Truths Break Similarity-Based Retrieval},
  author={Kargi, Bora and Uselis, Arnas and Oh, Seong Joon},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  url={https://arxiv.org/abs/2602.23906}
}
```