typhoon-ai/khammoon-data

khammoon data

Data used to build the typhoon-ai/khammoon Thai morpheme (คำมูล) segmenter, described in the technical report arXiv:2611.13012.

config	rows	labels	role
llm_labelled	147,999 BEST2009-train sentences + 32,207 WangchanThaiInstruct-train documents	LLM segmentation (Qwen3-30B-A3B) corrected in two rounds: (1) a convention table learnt from dev-train of the development set plus guideline Step-2 span merges; (2) Thai–Thai boundaries re-predicted by a model fine-tuned on dev-train (other boundaries kept from round 1)	training
dev	2,252 BEST2009 test sentences	human (3 annotators, adjudicated)	development: dev_split = devtrain (1,484; label correction + fine-tuning) / devval (768; model & threshold selection)
test	2,000 WangchanThaiInstruct test documents	human (3 annotators, adjudicated)	evaluation only — never used during development
llm_outputs	the dev and test texts	segmentation by the LLM segmenter that produced the LLM labels (Qwen3-30B-A3B-Instruct-2507, two-pass prompt v6-3, pass-1 T=0.2)	baseline: P 0.975 / R 0.948 / F1 0.962 on test
annotator_agreement	400 dev sentences + 400 test documents	the independent segmentation of each of the three annotators (tokens_a, tokens_b, tokens_c), before adjudication	inter-annotator agreement: Fleiss' κ 0.986 (BEST2009) / 0.991 (Wangchan)

The annotation guideline that defines a morpheme (คำมูล) for all labels in this dataset (human and LLM) is included as guideline/linguistic_guideline.md; its maintained copy lives in the code repository under linguistics/.

Every row satisfies "".join(tokens) == text; whitespace characters are kept as tokens. Every row has a license field (see below). llm_labelled rows also keep the original LLM segmentation in tokens_llm (before correction), so the whole pipeline can be rerun from this dataset (scripts/download_data.py in the code repository). test rows also have source_ids, the WangchanThaiInstruct ID(s) of the document (14 documents occur under more than one ID). Training rows whose text also appears in either human set were removed.

annotator_agreement rows hold the three annotators' independent segmentations of the same sentence, labelled A, B and C as in the technical report, plus reference_id, the id of the same sentence in dev or test whose tokens are the adjudicated segmentation. The annotation tool dropped newline characters, so text here can differ from the reference text in whitespace only; all three token lists join back to text exactly. scripts/annotator_agreement.py in the code repository recomputes the agreement figures of the report from this configuration.

Intended uses

This dataset is released in full so that developers and researchers can build on it:

• Train new segmenters. llm_labelled gives 70M characters of training text with both the corrected labels (tokens) and the original LLM labels (tokens_llm).
• Benchmark. dev and test are human-labelled; keep test for final evaluation only, as the khammoon results do, so numbers stay comparable.
• Study how native speakers disagree. annotator_agreement gives three independent segmentations of 800 sentences, the raw material behind the agreement figures.
• Compare with an LLM annotator. llm_outputs reproduces the strongest baseline without a GPU, and the difference between tokens_llm and tokens shows where the LLM goes wrong.
• Research Thai morphology and tokenization, e.g. morpheme-aware tokenizers for language models.

Evaluation

Recommended metric: boundary precision/recall/F1 ignoring boundaries adjacent to whitespace (see khammoon/metrics.py). khammoon 0.1.0 scores P 0.992 / R 0.993 / F1 0.992 on test.

License

The texts come from two sources and keep their licenses per row (license field). Because the sources mix NonCommercial and ShareAlike terms, there is no single license for the whole dataset; filter on license if you need a subset with specific terms.

rows	license	source and attribution
llm_labelled/best2009, dev	cc-by-nc-sa-3.0	BEST2009 corpus by NECTEC (nectec/best2009)
llm_labelled/wangchan (27,193 rows)	cc-by-nc-4.0	WangchanThaiInstruct by VISTEC/AIResearch (airesearch/WangchanThaiInstruct), per-row License of the source
llm_labelled/wangchan (5,014 rows)	cc-by-sa-4.0	same
test	cc-by-nc-4.0	WangchanThaiInstruct test split, same
annotator_agreement/best2009	cc-by-nc-sa-3.0	BEST2009, as dev
annotator_agreement/wangchan	cc-by-nc-4.0	WangchanThaiInstruct test split, as test

The segmentation labels added here (human labels in dev/test, corrected LLM labels in llm_labelled) are released under the same terms as the text of each row. None of the data may be used commercially except the cc-by-sa-4.0 rows. The LLM labels were produced with Qwen3-30B-A3B-Instruct-2507 (Apache-2.0).

Please cite BEST2009 / WangchanThaiInstruct when using the corresponding rows:

@inproceedings{limkonchotiwat2025thaiinstruct,
  title     = {WangchanThaiInstruct: An Instruction-Following Dataset for Culture-Aware, Multitask, and Multi-domain Evaluation in Thai},
  author    = {Limkonchotiwat, Peerat and Tuchinda, Pume and Lowphansirikul, Lalita and Nonesung, Surapon and Tasawong, Panuthep and Aji, Alham Fikri and Udomcharoenchaikit, Can and Nutanong, Sarana},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  year      = {2025},
  publisher = {Association for Computational Linguistics}
}