MIRACLE

MIRACLE is a multimodal benchmark dataset with image-based questions and model evaluation results.

This repository contains the test subset of the MIRACLE benchmark. The benchmark config provides the benchmark test split, and the model_results config provides test-set evaluation outputs for the included models.

Dataset Structure

The repository is organized as follows:

data/
  test.parquet          # Hugging Face loadable benchmark test split
  test.jsonl            # JSONL copy of the test split
  test.csv              # CSV copy of the test split
  Bench.tsv             # Original TSV metadata
  data.xlsx             # Original Excel metadata
images/
  batch1/
  batch2/
  batch3/
results/
  *_Bench_extract_result.xlsx
  test_results.parquet  # Long-format model test results
  test_results.jsonl

Fields

The benchmark config contains:

• index: sample id.
• question: question text.
• answer: ground-truth answer.
• image_path: list of image paths in this repository, prefixed with images/.
• source_image_path: original image paths from the source annotation file.
• type: question type.
• field: optional category field.
• num_images: number of images used by the sample.
• image_1 ... image_14: individual Hugging Face Image columns for Dataset Viewer previews. Empty image slots are null.

The model_results config contains the sample metadata and test result fields:

• model: model name parsed from the original result workbook filename.
• index: sample id.
• question: question text.
• answer: ground-truth answer.
• image_path: list of image paths in this repository, prefixed with images/.
• source_image_path: original image paths from the source annotation file.
• type: question type.
• field: optional category field.
• prediction: raw model prediction.
• extract_result: extracted final answer.
• score: evaluation score.

Loading

from datasets import load_dataset

bench = load_dataset("queyuecanyang/MIRACLE", "benchmark", split="test")
results = load_dataset("queyuecanyang/MIRACLE", "model_results", split="test")

Images are stored as files under images/. Each row's image_path column contains the relative paths for the images used by that sample. The image_1 ... image_14 columns contain the same files as Hugging Face Image features so that Dataset Viewer can render the actual images.