jmj-minju/transformers-en-ko-aligned-docs

Transformers EN-KO Aligned Docs

This dataset package contains English-Korean aligned text pairs derived from the docs/source/en and docs/source/ko trees in huggingface/transformers.

Repository layout

• data/: published dataset splits only
• metadata/: filtering, blacklist, and build-status artifacts
• docs/: agent harness and dataset construction notes
• AGENTS.md: short Codex entry point for this dataset repo

Contents

• data/train.jsonl: final training split with manually reviewed blacklist applied
• data/validation_clean.jsonl: clean validation split sampled by document from the final training set
• data/stress_test_contaminated.jsonl: document pairs excluded from training because they contained alignment or contamination issues
• metadata/blacklist.txt: documents manually excluded after review
• metadata/doc_split.csv: document-level clean vs contaminated split
• metadata/rejected_docs.csv: rejected-document summary with reasons
• metadata/build_status.json: machine-readable build summary
• docs/translation_dataset_harness.md: detailed harness and review policy
• docs/dataset_methodology.md: concise build and usage notes

Row counts

• train rows: 76
• validation rows: 5
• stress-test rows: 880
• blacklist docs: 13

Schema

Each JSONL row contains:

• relative_path
• chunk_id
• text
• target
• source_chars
• target_chars

Build notes

• source repository: huggingface/transformers
• source commit used locally: ce7efd8f19
• alignment method: markdown normalization, one-sided metadata removal, conservative block-level alignment, document rejection for structural mismatch, and manual blacklist
• preprocessing note: the current exported dataset was built with an earlier cleanup-heavy pipeline, but the harness now recommends preserving headings and code-block structure during alignment and cleaning more aggressively only at export time

Intended use

• train.jsonl is the conservative split intended for fine-tuning.
• validation_clean.jsonl is the clean validation split for model selection.
• stress_test_contaminated.jsonl is not a gold benchmark. It is a stress-test set containing documents that were rejected from training because alignment looked unreliable.

Limitations

• The dataset is partially automatically aligned and then manually filtered.
• Some original docs are localized or rewritten rather than directly translated, and those files should not be treated as parallel data.
• Review the blacklist and rejected reports before using this dataset for evaluation claims.

Source and attribution

Original documentation content comes from the Hugging Face Transformers repository:

• English: docs/source/en
• Korean: docs/source/ko

Before publishing publicly, verify that your redistribution plan matches the upstream repository license and your intended use.