MyanParemia Initial Dataset

I built MyanParemia as a human-validated dataset and reproducible processing pipeline for Burmese sayings and proverbs, digitized from two Myanmar Language Commission books. This snapshot is the copyright-safe structural metadata view of that work.

The initial snapshot contains metadata for 1,953 validated entries:

Collection	Source	Entries	Index links	Review status
Sayings	Myanmar Language Commission, 1996	1,010	3,182	All valid
Proverbs	Myanmar Language Commission, 2019	943	2,864	All valid
Total		1,953	6,046	All valid

What I am publishing in this initial snapshot

data/entries.jsonl has one row per source-book entry: stable identifiers, collection membership, validation status, source metadata, and structural field counts. It is a real, loadable dataset view generated from my gold records, but I deliberately exclude the raw OCR, headword text, meanings, quotations, origins, and index terms.

I am taking this staged approach because the source books are copyrighted. Publishing this structural view does not mean I am claiming the source transcription is openly licensed. I will confirm permission and redistribution terms before a wider, text-bearing release.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset("kaunghsethein/myanparemia", "metadata")
print(dataset["train"][0])

To inspect the folder locally without an account:

from datasets import load_dataset

dataset = load_dataset("json", data_files="data/entries.jsonl")

How I reproduce the public snapshot

From my main MyanParemia development repository:

uv run python report/huggingface/scripts/build_public_snapshot.py

The builder reads data/gold/saying.jsonl and data/gold/proverb.jsonl, validates the required identity and review fields, and regenerates both the public JSONL and dataset_summary.json.

Local research view

If you lawfully possess the source books, you can build a local, non-public research view:

uv run python report/huggingface/scripts/prepare_local_research_view.py \
  --gold-dir data/gold \
  --output data/research-local.jsonl

The generated research-local.jsonl is git-ignored in my repository. I do not upload it, and you should not redistribute it, unless redistribution permission has been established.

Pipeline provenance

My development pipeline keeps three layers:

1. raw OCR and deterministic entry boundaries;
2. structured, provisional JSONL;
3. human-reviewed gold JSONL with an append-only validation log.

Extraction preserves Burmese text and the punctuation marks ။ and ၊ verbatim. Normalization, index joining, and benchmark construction are separate passes. Every entry uses a stable key such as saying:0001 or proverb:0001 in addition to its book-local printed number.

My initial benchmark work includes a reproducible 12-item meaning-selection pilot with a fixed seed, stored prompts, raw model outputs, and strict answer parsing. Its distractors still need task-level human review, so I am not presenting its scores as final benchmark results.

Intended uses

• auditing collection coverage and structural complexity;
• developing a permission-aware Burmese dataset release;
• reproducing my extraction and validation pipeline;
• preparing human-reviewed figurative-language benchmark tasks.

Limitations

• This public snapshot contains structural metadata, not the complete text-bearing gold records.
• I checked OCR correction and structural classification against the printed sources, but I do not modernize or normalize the source language.
• Raw index terms are preserved in my private gold layer and are not yet classified into a final semantic taxonomy.
• The meaning-selection task is a small pipeline demonstration, not a reportable benchmark.

Citation

I have not assigned a formal paper citation or permanent release identifier yet. Until then, please cite the repository name, snapshot date, and commit hash you used.

Contact

Kaung Hset Hein, Innovative Cognitive Computing Research Center (IC2), King Mongkut's University of Technology Thonburi (KMUTT). kaunghset.hein@mail.kmutt.ac.th