ECA-ToS-Benchmarks

This dataset provides the ToS benchmark annotations used in ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation.

ToS is designed for continual learning in open-ended image-to-text generation. Each image is assigned to a task by its dominant visual topic. Other visible topics remain in the sample, so tasks shift over time while shared concepts can still reappear across tasks.

Dataset Contents

This repository contains annotation files and topic metadata for four benchmarks.

Benchmark	Source dataset	Task	Annotation files
ToS-COCO Caption	MSCOCO Caption	Image Captioning	annotations/coco/tos_coco_caption_*.json
ToS-VQAv2	VQAv2	Visual Question Answering	annotations/coco/tos_vqav2_*.json
ToS-TextCaps	TextCaps	Image Captioning	annotations/text/tos_textcaps_caption_*.json
ToS-TextVQA	TextVQA	Visual Question Answering	annotations/text/tos_textvqa_*.json

The topic metadata files record the dominant topic and topic composition used to build the splits.

File	Description
annotations/coco/tos_coco_style_topic_metadata.json	Topic metadata for COCO-style splits
annotations/text/tos_text_style_topic_metadata.json	Topic metadata shared by ToS-TextCaps and ToS-TextVQA

The repository also includes the official VQAv2 validation files required by the VQA evaluator.

annotations/coco/answer_list.json
annotations/coco/v2_OpenEnded_mscoco_val2014_questions.json
annotations/coco/v2_mscoco_val2014_annotations.json

Images

Images are not redistributed in this dataset repository.

Please download images from the original dataset sources and place them in the layout expected by the ECA codebase.

cache/coco/images/
cache/TextCaps/images/

COCO-style benchmarks use MSCOCO 2014 train and validation images. Text-style benchmarks use the TextCaps and TextVQA image folders.

Using With The ECA Codebase

Download this dataset and copy the annotation files into the local cache/ directory.

huggingface-cli download Snowball0823/ECA-ToS-Benchmarks   --repo-type dataset   --local-dir data/ECA-ToS-Benchmarks   --local-dir-use-symlinks False

mkdir -p cache/coco/annotations cache/TextCaps
cp data/ECA-ToS-Benchmarks/annotations/coco/*.json cache/coco/annotations/
cp data/ECA-ToS-Benchmarks/annotations/text/*.json cache/TextCaps/

After copying, the expected runtime layout is:

cache/
  coco/
    annotations/
      tos_coco_caption_train.json
      tos_coco_caption_val.json
      tos_coco_caption_test.json
      tos_vqav2_train.json
      tos_vqav2_val_eval.json
      tos_coco_style_topic_metadata.json
      answer_list.json
      v2_OpenEnded_mscoco_val2014_questions.json
      v2_mscoco_val2014_annotations.json
    images/
      train2014/
      val2014/
  TextCaps/
    tos_textcaps_caption_train.json
    tos_textcaps_caption_val.json
    tos_textcaps_caption_val_eval.json
    tos_textvqa_train.json
    tos_textvqa_val.json
    tos_textvqa_val_eval.json
    tos_text_style_topic_metadata.json
    images/
      train/
      test/

License And Source Datasets

This dataset is released under CC BY 4.0. The annotations are derived from MSCOCO Caption, VQAv2, TextCaps, and TextVQA. The original images and annotations remain subject to the licenses and terms of their source datasets.

Users should also cite and follow the original dataset sources.

• COCO: https://cocodataset.org/
• VQAv2: https://visualqa.org/
• TextVQA and TextCaps: https://textvqa.org/

Citation

If you use this dataset, please cite our paper.

@inproceedings{kong2026eca,
  title={ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation},
  author={Kong, Jiangtao and Zhao, Peijun and Chen, Chun-Fu and Do, Youngwook and Hu, Shaohan and Zhou, Tianyi and Shao, Huajie},
  booktitle={International Conference on Machine Learning},
  year={2026}
}