Hugging Face's logo

Datasets:
Davidsamuel101
/
MUDIDI

Tasks:
Text Generation
Modalities:
Document
Text
Formats:
parquet
Languages:
Multiple languages
Size:
< 1K
ArXiv:
Tags:
dictionary-digitization
ocr
multilingual
mdf
mudidi
lexicography
DOI:
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1

Protecting the integrity of the MUDIDI benchmark for evaluation

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

MUDIDI is a gated evaluation benchmark for multilingual dictionary digitization. Gold annotations are released under CC BY-NC 4.0; source page images are derived from public-domain HathiTrust scans. By requesting access you agree to the conditions below.

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

MUDIDI Dataset

Gold annotations and source pages for MUDIDI, a two-stage benchmark for multilingual dictionary digitization.

Access: This dataset is gated (gated: auto). Log in, open the dataset page, accept the benchmark access conditions (non-commercial use, no crawler-exposed re-hosting, no training-data contamination when evaluating), then authenticate with a Hugging Face token before downloading.

Dataset summary

Property Value
Dictionaries 30 public-domain multilingual dictionaries (26 bilingual; 3 trilingual; 1 four-language: Na-English-Chinese-French)
Stage 1 gold pages 90 annotated pages across 30 dictionaries (3 per dictionary)
Stage 1 source pages 90 source page files (75 PDFs; 15 PNGs; 3 per dictionary; every page has Stage 1 TSV, flat transcript, and language map)
Stage 2 pages 10 (1 representative page per Stage 2 subset dictionary)
Writing systems Latin, Cyrillic, Greek, Devanagari, Bengali, Gujarati, Gurmukhi, Telugu, Hebrew, Syriac, Arabic-based, Khmer, Han, Kana, Cuneiform, IPA, and more
Splits One train split per configuration; no train/test partition (evaluation benchmark only)
Version v1.0.0

Repository layout

Top level:

.
├── .gitattributes
├── .gitignore
├── LICENSE
├── README.md
├── build_parquet.py
├── parquet/
│   └── <config>/train.parquet
└── dictionaries/
    └── <Source-Target>/
        ├── dictionary_languages.yaml
        ├── Alphabet list/
        ├── Dictionary pages/
        ├── Stage 1 Gold OCR/
        ├── Stage 2 Gold Cheat Sheet/   # 10 dictionaries only
        └── Stage 2 MDF file/            # 10 dictionaries only

Each dictionary lives under dictionaries/<Source-Target>/:

Path Contents Format
dictionary_languages.yaml Source/target language codes; writing-system and HathiTrust metadata when available YAML
Alphabet list/ Source-language character or orthography inventory (optional hint for Stage 1) alphabet.txt with dictionary-specific entries
Dictionary pages/ Three sampled dictionary entry pages page_<N>.png or page_<N>.pdf
Stage 1 Gold OCR/ Human column-aligned gold, derived flat transcript, and language-span map page_<N>/page_<N>_stage1_GOLD.tsv, page_<N>_stage1_GOLD_flat.txt, page_<N>_lang.json
Stage 2 Gold Cheat Sheet/ Human MDF marker schema (10 dictionaries only) mdf_parsing_guide.json
Stage 2 MDF file/ Human Toolbox MDF records and language-projection provenance (10 dictionaries only) page_<N>.mdf.txt, page_<N>_mdf_lang_projection.json

Example (Evenki-Russian):

dictionaries/Evenki-Russian/
├── dictionary_languages.yaml
├── Alphabet list/alphabet.txt
├── Dictionary pages/page_1.png
├── Dictionary pages/page_2.png
├── Dictionary pages/page_3.png
├── Stage 1 Gold OCR/page_1/page_1_stage1_GOLD.tsv
├── Stage 1 Gold OCR/page_1/page_1_stage1_GOLD_flat.txt
├── Stage 1 Gold OCR/page_1/page_1_lang.json
├── Stage 2 Gold Cheat Sheet/mdf_parsing_guide.json
├── Stage 2 MDF file/page_1/page_1.mdf.txt
└── Stage 2 MDF file/page_1/page_1_mdf_lang_projection.json

No Introduction/ folders are included in this release.

Stage 1 gold format

Gold files live under Stage 1 Gold OCR/page_<N>/:

  • page_<N>_stage1_GOLD.tsv — column-aligned human annotation (primary gold).
  • page_<N>_stage1_GOLD_flat.txt — derived flat transcript: one line per visible row, column-major reading order, with <b> / <i> markup preserved.
  • page_<N>_lang.json — character-level language/script spans bound to the exact flat transcript.

The three Stage 1 artifacts are present for all 90 sampled pages. The Circassian trilingual TSV stores separate aligned column cells, while its flat transcript joins aligned columns with |; its flat-line count therefore does not equal its TSV-cell count.

The Stage 1 artifact details are described in the Stage 1 gold format section above.

Stage 2 gold format

Stage 2 gold is available for 10 dictionaries only:

  • Evenki-Russian, Chukchi-Russian, Nahuatl-French, Na-English-Chinese-French, Kashmiri-English, Tiri-English, Greek-English, Efik-English, Circassian-English-Turkish, Iñupiatun Eskimo-English

  • mdf_parsing_guide.json — maps dictionary-specific MDF markers to entry structure rules (Pass 1 gold).

  • page_<N>.mdf.txt — blank-line-delimited SIL Toolbox MDF records (Pass 2 gold).

  • page_<N>_mdf_lang_projection.json — field-level provenance from MDF values to Stage 1 language/script spans; paths are dataset-root-relative and hashes bind the referenced files.

The Stage 2 artifact details are described in the Stage 2 gold format section above.

Loading with datasets

Tabular gold text is exported under parquet/<config>/train.parquet (one row per annotated page). Each configuration maps to one multilingual dictionary and exposes a single train split with:

Column Source
page_id Page stem, e.g. page_54
ocr_text Stage 1 Gold OCR/page_<N>/page_<N>_stage1_GOLD_flat.txt
mdf_text Stage 2 MDF file/page_<N>/page_<N>.mdf.txt when present, else empty 
from datasets import load_dataset

ds = load_dataset("Davidsamuel101/MUDIDI", "bengalese-english", split="train")
print(ds[0]["page_id"])
print(ds[0]["ocr_text"][:200])
print(ds[0]["mdf_text"])  # empty unless this page has Stage 2 MDF gold

Install the dataset tooling with uv sync --extra dataset. To regenerate Parquet from the source checkout after editing gold files:

uv run --extra dataset python scripts/build_dataset_parquet.py

After downloading the HF repository, run its root copy with:

uv run --extra dataset python build_parquet.py

Downloading the dataset

Hugging Face (gated)

  1. Open huggingface.co/datasets/Davidsamuel101/MUDIDI while logged in and accept the dataset terms.
  2. Create a token with read access at huggingface.co/settings/tokens.
  3. Authenticate locally:
hf auth login
# or: export HF_TOKEN=hf_...
  1. Download the full dataset:
# into the MUDIDI repo layout used by examples/
hf download Davidsamuel101/MUDIDI --repo-type dataset --local-dir dataset/MUDIDI

For a custom destination:

hf download Davidsamuel101/MUDIDI --repo-type dataset --local-dir /path/to/mudidi

Large downloads can also use:

hf download Davidsamuel101/MUDIDI --repo-type dataset --local-dir dataset/MUDIDI --max-workers 8

Raw files (without datasets) 

from pathlib import Path

from huggingface_hub import login, snapshot_download

login()  # or set HF_TOKEN
root = Path(snapshot_download("Davidsamuel101/MUDIDI", repo_type="dataset"))

# Stage 1 evaluation pages (90 gold TSV files)
stage1_pages = sorted(root.glob("dictionaries/*/Stage 1 Gold OCR/page_*/*_stage1_GOLD.tsv"))

# Stage 2 dictionaries (10 with MDF gold)
stage2_dicts = sorted(
    p.parents[1].name
    for p in root.glob("dictionaries/*/Stage 2 Gold Cheat Sheet/mdf_parsing_guide.json")
)

# Example: Evenki-Russian page_1 paths
evenki = root / "dictionaries/Evenki-Russian"
page_image = evenki / "Dictionary pages/page_1.png"
page_gold = evenki / "Stage 1 Gold OCR/page_1/page_1_stage1_GOLD_flat.txt"
page_lang = evenki / "Stage 1 Gold OCR/page_1/page_1_lang.json"
page_mdf = evenki / "Stage 2 MDF file/page_1/page_1.mdf.txt"
page_projection = evenki / "Stage 2 MDF file/page_1/page_1_mdf_lang_projection.json"

Source repository

The dataset files are published in the Hugging Face dataset repository and are not vendored in the GitHub source checkout. The source checkout contains the tracked scripts/build_dataset_parquet.py; the HF copy is uploaded as build_parquet.py at the dataset root.

Known limitations

  • Stage 2 gold covers 10 of 30 dictionaries; do not assume MDF gold for all languages.
  • All 90 sampled pages have Stage 1 TSV, flat-transcript, and language-map gold artifacts.
  • Introduction/ folders are not included in this release.
  • Source files are public-domain-derived HathiTrust scans in PDF or PNG form; typography and scan quality vary.

Licensing

  • Annotations (gold TSV, flat transcripts, language-span maps, MDF records, parsing guides, projections, and alphabet lists): CC BY-NC 4.0
  • Source page images: derived from public-domain HathiTrust volumes; consult the original volume for any residual usage constraints.

See LICENSE for details.

Citation

If you use the dataset:

@misc{mudidi_v1,
  title        = {{MUDIDI: Multilingual Dictionary Digitization Benchmark}},
  author       = {Setiawan, David and Khishigsuren, Temuulen and Agarwal, Milind and Pit, Pagnarith and Mahmudi, Aso and Vylomova, Ekaterina},
  year         = {2026},
  version      = {v1.0.0},
  howpublished = {\url{https://huggingface.co/datasets/Davidsamuel101/MUDIDI}}
}

If you use the framework or report benchmark results from the paper:

@misc{mudidi2026,
  title         = {{MUDIDI: A Two-Stage Framework for Multilingual Dictionary Digitization with Language Models}},
  author        = {Setiawan, David and Khishigsuren, Temuulen and Agarwal, Milind and Pit, Pagnarith and Mahmudi, Aso and Vylomova, Ekaterina},
  year          = {2026},
  eprint        = {2606.09435},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2606.09435},
  url           = {https://arxiv.org/abs/2606.09435},
  note          = {Submitted to EMNLP 2026}
}
Downloads last month
46
Total file size:
74.2 MB

Paper for Davidsamuel101/MUDIDI