Datasets:

Modalities:
Text
Formats:
json
Languages:
Thai
ArXiv:
License:
File size: 2,691 Bytes
b35eaa3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
OpenThai2.0-WordSeg-Benchmark

Code (*.py): Apache License 2.0 (LICENSE).

The test files in data/ are built from the sources below. Each keeps the licence of its source, and the files
built from a CC BY-SA source are shared under the same licence. All credit for the sources goes to their authors.

- tud.jsonl: UD_Thai-TUD, test split (Sriwirote et al., 2025, The Thai Universal Dependency Treebank, TACL 13).
  Universal Dependencies, revision d4cb5fd3. https://github.com/UniversalDependencies/UD_Thai-TUD
  Licence: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/
- pud.jsonl, pudA.jsonl, pudB.jsonl: UD_Thai-PUD (Zeman et al., 2017, CoNLL 2017 Shared Task).
  Universal Dependencies, revision 4e0e10fd. https://github.com/UniversalDependencies/UD_Thai-PUD
  Licence: CC BY-SA 3.0, https://creativecommons.org/licenses/by-sa/3.0/
- ws160.jsonl: Wisesight-160 word segmentation samples (Suriyawongkul et al., 2024, PyThaiNLP/Wisesight
  Sentiment Corpus with Word Tokenization Label, version 1.1).
  PyThaiNLP, revision 6995f993. https://github.com/PyThaiNLP/wisesight-sentiment
  Licence: CC0 1.0, https://creativecommons.org/publicdomain/zero/1.0/
- vistec.jsonl: VISTEC-TP-TH-2021, test split (Limkonchotiwat et al., 2021, Handling Cross- and Out-of-Domain
  Samples in Thai Word Segmentation, Findings of ACL-IJCNLP 2021), from the OSKut repository, revision be9e16ac.
  https://github.com/mrpeerat/OSKut
  Licence: CC BY-SA 3.0, https://creativecommons.org/licenses/by-sa/3.0/

Changes made to these sources (build_testsets.py): each file is converted to natural text, one sentence or post
per line, with the character offsets of its gold words; a space always separates two words, so a source token
with a space inside is split; VISTEC's markup is removed; pudA and pudB are the two halves of the PUD test set.
Defects of the sources are kept as published.

Not redistributed (non-commercial licences); build_testsets.py rebuilds them from their original sources:
- BEST 2009 (Kosawat et al., 2009), NECTEC, CC BY-NC-SA 3.0. https://huggingface.co/datasets/nectec/best2009
- ORCHID (Sornlertlamvanich et al., 1999), NECTEC, CC BY-NC-SA 3.0.
  https://github.com/wannaphong/corpus_mirror/releases/tag/orchid-v1.0
- khammoon test set (Sirichotedumrong et al., 2026), Typhoon, CC BY-NC 4.0; its documents come from
  WangchanThaiInstruct (Limkonchotiwat et al., 2025). https://huggingface.co/datasets/typhoon-ai/khammoon-data

predictions/: the outputs of the systems listed in README.md. The word files hold line ids and character offsets
only, no text. The token files exist only for tud, pud, ws160 and vistec; they contain the text of these sets,
under its licence.