You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ExpertHTR Dataset

Gated page-level handwritten text recognition data for the ExpertHTR project.

This is a rights-filtered replacement export: all HWDB/CASIA records and images have been removed. The repository remains gated because the remaining upstream sources have different access conditions. It is a companion data release for ExpertHTR, not the exact training snapshot for the published seven-source checkpoint.

Included data

Split Records Purpose
train 14,875 Model training
valid 662 Checkpoint selection
test 912 Final evaluation

The records support four tasks: page_htr, line_count, phrase_line_lookup, and line_range_htr.

Files and schema

page_multitask/train.jsonl
page_multitask/valid.jsonl
page_multitask/test.jsonl
img/                         # images referenced by `images`
RELEASE_MANIFEST.json
README.md

Each JSONL row contains id, an images list with root-relative paths, multimodal messages, text_raw, and a deliberately minimal metadata object. The public metadata keeps only source_dataset, task, split, language, and script; writer identifiers, source filenames, raw annotation provenance, bboxes and archive hashes were removed. Resolve images[0] from the dataset root.

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("DAIR-Group/ExpertHTR-Dataset", "page_multitask", token=True)
print(ds)
sample = ds["test"][0]
print(sample["id"], sample["images"][0])

Accept the gated access terms before downloading. For a local copy, replace the repository name with the directory containing page_multitask/.

Access and licensing

This is a gated research release. The collection is derived from multiple upstream datasets; each source retains its own copyright and license. Do not redistribute, sell, or use commercially without checking the upstream terms and obtaining required permission.

HWDB/CASIA is intentionally absent. Obtain it directly from its official provider if needed and follow its application and research-use conditions.

RELEASE_MANIFEST.json records the exclusion policy and split counts. See the ExpertHTR code repository for audits, training and evaluation commands.

Security and privacy

The release contains no API keys, access tokens, email addresses or local filesystem paths. Do not add credentials to JSONL metadata, commit messages or Space variables. Keep the repository gated and review every future file before making it public.

Downloads last month
363