tinnel123/defacto_1.5k_benchmark

DeFacto Subset Dataset Notes

This directory contains a compact subset for evaluating evidence alignment and faithful reasoning in multimodal tasks. The data format follows the ideas in "DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning" (arXiv:2509.20912), emphasizing both answer correctness and evidence-region alignment.

Directory Structure

• filtered_100_per_file.json: main annotation file
• images/: image directory (samples reference via relative paths)
• upload.py: optional Hugging Face dataset upload script

Data Format

Each sample is a JSON object with the following fields:

• id: unique sample identifier
• conversations: a list of messages from user/assistant
• images: list of image paths for the sample (relative to images/)

The assistant output uses the following tags:

• : list of evidence rectangles, format [[x1, y1, x2, y2], ...]
• : final answer
• : intermediate reasoning (optional, for training/analysis; not recommended for public evaluation)

Example

{
    "id": "data/DUE_Benchmark/KleisterCharity/page_pngs/abbf98ed31e28068150dce58296302ee_page0.png_1",
    "conversations": [
        {"from": "user", "value": "What is the value for the address__post_town?\n<image>"},
        {"from": "assistant", "value": "<bbox>[[153, 268, 829, 317], [374, 250, 605, 264]]</bbox>\n<think>...</think>\n<answer>HAMPTON</answer>"}
    ],
    "images": ["images/abbf98ed31e28068150dce58296302ee_page0.png"]
}

Usage

1. Load filtered_100_per_file.json
2. Load images using samples[i].images
3. Parse and in the assistant message for evidence-alignment evaluation

Notes

• coordinates are in original image pixel space.
• If the model answers correctly but the evidence region is irrelevant, it is a typical case of "correct answer but unfaithful reasoning," which matches the paper's focus.

Citation

If you use this data or the idea, please cite:

• Tianrun Xu et al. DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning. arXiv:2509.20912.
• or

@misc{xu2026defactocounterfactualthinkingimages,
     title={DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning}, 
     author={Tianrun Xu and Haoda Jing and Ye Li and Yuquan Wei and Jun Feng and Guanyu Chen and Haichuan Gao and Tianren Zhang and Feng Chen},
     year={2026},
     eprint={2509.20912},
     archivePrefix={arXiv},
     primaryClass={cs.AI},
     url={https://arxiv.org/abs/2509.20912}, 
}