hmhm1229/ConceptFormer-Trainset

ConceptFormer Trainset

Grounded visual-document retrieval training data for ConceptFormer. The release contains 37,966 training samples and an image-complete, deduplicated corpus of 12,232 positive document images. Multi-image relations are preserved through each sample's relevant_doc_ids list.

from datasets import load_dataset

annotations = load_dataset(
    "parquet",
    data_files="hf://datasets/hmhm1229/ConceptFormer-Trainset/data/train.parquet",
    split="train",
)
corpus = load_dataset(
    "parquet",
    data_files="hf://datasets/hmhm1229/ConceptFormer-Trainset/data/corpus-*.parquet",
    split="train",
)

Each row contains query_id, query_text, relevant_doc_ids, describe, bbox_2d, image_size, input_size, answers, and dataset_names. Bounding boxes are stored as JSON-encoded xyxy coordinates. See the repository docs/DATA.md and docs/ANNOTATION.md for schema and construction details.

The annotation table has 47,389 positive-document references: 28,547 single-image samples, 9,417 two-image samples, one three-image sample, and one five-image sample. Corpus images are stored once as doc_id, image: {bytes, path}, and dataset_name.

Paper can be seen in Arxiv

License and Provenance

Derived annotations are released for research use. Images and source questions retain the licenses and terms of their upstream datasets. Users are responsible for complying with those terms.