VL-CheckList

This repository provides a Hugging Face format of VL-CheckList, an evaluation framework for Vision-Language Models (VLMs) and Vision-Language Pre-training (VLP).

Instead of relying solely on coarse downstream task metrics, VL-CheckList isolates and diagnoses model capabilities across three fundamental visual-linguistic dimensions: Object, Attribute, and Relation.

• Official Paper: VL-CheckList: Evaluating Pre-trained Vision-Language Models with Objects, Attributes and Relations (EMNLP 2022 System Demonstrations)
• Original GitHub: om-ai-lab/VL-CheckList
• Original Demo: Hugging Face Space Demo

---

Benchmark Overview

All 55 fine-grained sub-tasks from the original paper have been consolidated into 3 unified evaluation splits (~1.08M test pairs) alongside a fast image lookup table:

Split / Config	Samples	Description	Evaluated Sub-tasks
object	733,433	Entity existence, bounding area, and positioning	Location (center, mid, margin), Size (small, medium, large)
relation	244,831	Interactions and geometric arrangements	Action, Spatial relationships
attribute	110,627	Fine-grained object descriptors	Color, Material, Shape, State
image_index	177,733	Fast $O(1)$ lookup index mapping image_id to image shards	VG, HAKE (2019/2020), SWiG, HICO, OpenImages, V-COCO

---

Data Schema

Each sample in object, attribute, and relation represents a contrastive test case:

{
  "image_id": "HOI_PVP_00000004.jpg",
  "image_path": "hake_images_20190730/HOI_PVP_00000004.jpg",
  "pos_item": ["person hose dog"],     # Ground truth positive caption/label
  "neg_item": ["person hose pizza"],   # Contrastive foil / negative caption
  "sub_task": "location_hake_obj_center",
  "axis": "object",
  "image_subset": "images_hake_2019",  # Image shard directory
  "shard_file": "data-00000-of-00003.parquet"
}

Quickstart: Loading Data & Images

1. Load Benchmark Labels

from datasets import load_dataset

# Load a full axis (e.g., 'attribute')
dataset = load_dataset("w4nn4b3M4ST3R/vl-checklist", "attribute", split="test")

# Optional: filter by specific sub-task
color_tests = dataset.filter(lambda x: x["sub_task"] == "vaw_color")
print(color_tests[0])

2. Fetch Corresponding Images

The images are decoupled from the test labels to prevent sample duplication. You can resolve images on the fly:

import io
from PIL import Image
import pyarrow.parquet as pq
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
REPO_ID = "w4nn4b3M4ST3R/vl-checklist"

def fetch_image(sample, repo_id=REPO_ID):
    """Fetches and decodes the PIL Image for a given test sample."""
    subset = sample["image_subset"]
    shard = sample["shard_file"]
    img_id = sample["image_id"]
    
    file_path = f"datasets/{repo_id}/data/{subset}/{shard}"
    with fs.open(file_path, "rb") as f:
        table = pq.read_table(f, columns=["image_id", "bytes"]).to_pandas()
        row = table[table["image_id"] == img_id]
        if not row.empty:
            return Image.open(io.BytesIO(row.iloc[0]["bytes"])).convert("RGB")
    return None

# Example usage:
sample = color_tests[0]
image = fetch_image(sample)
print(f"Loaded image: {sample['image_id']} with size {image.size}")

Citation

If you use this benchmark in your research, please cite the original authors:

@inproceedings{zhao2022explainable,
  title={An explainable toolbox for evaluating pre-trained vision-language models},
  author={Zhao, Tiancheng and Zhang, Tianqi and Zhu, Mingwei and Shen, Haozhan and Lee, Kyusong and Lu, Xiaopeng and Yin, Jianwei},
  booktitle={Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
  pages={30--37},
  year={2022}
}

@article{zhao2022vlchecklist,
  title={VL-CheckList: Evaluating Pre-trained Vision-Language Models with Objects, Attributes and Relations},
  author={Zhao, Tiancheng and Zhang, Tianqi and Zhu, Mingwei and Shen, Haozhan and Lee, Kyusong and Lu, Xiaopeng and Yin, Jianwei},
  journal={arXiv preprint arXiv:2207.00221},
  year={2022}
}