CS-CLIP-data / README.md
juns's picture
Release CS-CLIP training data and benchmark files
d224bee verified
|
Raw History Blame Contribute Delete
4.25 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: CS-CLIP data
size_categories:
  - 1M<n<10M
task_categories:
  - zero-shot-image-classification
tags:
  - compositional-reasoning
  - hard-negatives
  - clip
configs:
  - config_name: train
    data_files: train/cs_clip_hard_negatives.jsonl

CS-CLIP data

Training data and benchmark files for CS-CLIP: Compositional Scene Graph-guided CLIP for Robust Compositional Reasoning (Findings of EMNLP 2026).

Images are not included. Download COCO, Visual Genome (ARO) and GQA images from their original sources; the code repository explains how.

Files

File Records Content
train/cs_clip_hard_negatives.jsonl 1,941,340 Caption / hard-negative pairs on 78,614 COCO train2014 images. This is the exact training set of the released model.
eval/gqa_two_choice.json 10,000 GQA two-choice benchmark built from the GQA balanced validation split.
eval/controlled_object.json 23,937 Controlled test cases: one object replaced (from VG-Relation).
eval/controlled_attribute.json 28,748 Controlled test cases: one attribute replaced (from VG-Attribute).
eval/controlled_relation.json 13,595 Controlled test cases: one relation replaced (from VG-Relation).

Training data

Each line is one hard negative:

{"image_id": 318556, "image_path": "train2014/COCO_train2014_000000318556.jpg",
 "caption": "A very clean and well decorated empty bathroom",
 "negative_caption": "A very clean and well decorated empty bed", "type": "REPLACE-OBJ"}
  • caption is a human-written COCO caption.
  • negative_caption changes one compositional element of the caption.
  • type gives the operation:
Type Count
REPLACE-OBJ 354,906
REPLACE-ATT 174,368
REPLACE-REL 311,156
SWAP-OBJ 318,529
SWAP-ATT 41,821
ADD-OBJ 365,583
ADD-ATT 374,977

How the negatives were made:

  1. Captions are parsed into scene graphs with SceneGraphParser.
  2. For Replace and Add, bert-base-uncased proposes candidates for the target span, and roberta-large-mnli keeps the most contradictory one.
  3. Swap is rule-based.

The negatives are generated automatically and are not human-verified.

  • A small fraction is ungrammatical, or differs from the caption only in letter case (about 4K SWAP-OBJ items).
  • Because the masked language model fills words freely, some negatives contain violent, sexual or profane words that are not in the original caption, for example "dead", "corpse", "naked" or "prostitute". This affects about 0.2% of pairs.
  • The negatives are meant as contrastive training signal. Do not use them as natural image descriptions.

Benchmark files

  • GQA two-choice. {"question_id", "image", "question", "positive", "negative"}. positive is the question followed by the gold full answer; negative replaces the answer with another entity from the same image's scene graph. image is relative to the GQA image folder.
  • Controlled test cases. {"image", "positive", "negative"}. They are built from the VG-Relation and VG-Attribute sets of ARO (Yuksekgonul et al., ICLR 2023). Each negative replaces exactly one object, attribute or relation of the positive caption with a word sampled from the Visual Genome vocabulary. image is relative to the ARO image folder (vgr_vga_images.zip). Models are evaluated on full images.

License

CC BY 4.0. Captions come from COCO and annotations from Visual Genome and GQA, all released under CC BY 4.0. The controlled test cases are derived from ARO's VG-Relation and VG-Attribute. Images are not included and remain subject to their original terms.

Citation

@inproceedings{jeong2026csclip,
  title     = {{CS-CLIP}: Compositional Scene Graph-guided {CLIP} for Robust Compositional Reasoning},
  author    = {Jeong, SeongJun and Jung, Minjoon and Choi, Woo Suk and Jang, Youwon and Zhang, Byoung-Tak},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}