haznitrama/idn-ban-cbn-synthetic

Indonesian–Balinese–Cirebonese Synthetic Parallel Corpus

Dataset Summary

This corpus is fully synthetic and targets two extremely low-resource languages:

• Balinese (ban)
• Cirebonese (often grouped under Javanese dialects)

Key facts:

• Seeded with 10k high-level topics generated by GPT-5.
• For each topic, gpt-oss-120b produced 30 subtopics and continued to generate 30 question–answer pairs (≈9M Q&A pairs).
• Answers were translated into Balinese and Cirebonese by gpt-oss-120b while leveraging bilingual lexicons/dictionaries through prompting and in-context learning.
• The parallel corpus contains roughly ~1B tokens per low-resource language (Balinese & Cirebonese) across all splits.

Available Subsets

• raw: direct aggregation of every translated answer that passes structural validation (IDs and all three language fields present).
• filtered_heuristic: subset of raw that passes filtering based on heuristics including minimum length, repetition checks, and also GlotLID verification for Balinese and Cirebonese (cirebonese are referred to as "jav" in GlotLID).

Citation

If you use this dataset (or derivatives), please cite:

@misc{scale-lowres-synth-2025,
  title        = {Scale Resources for Low-Resource Languages via Synthetic Data Generation},
  author       = {Faiz Ghifari Haznitrama and Najma Qalbi Dwiharani and Alice Oh},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/haznitrama/idn-ban-cbn-synthetic}},
}

Notes

• Dataset is synthetic; no human-written or human-translated text is included.
• Licensed under the permissive MIT terms to encourage downstream reuse.
• Please verify downstream safety/quality constraints that apply to your deployment scenario before production use.