ChindaMT-Grounded

Paper: Reference-Grounded Data Curation for Instruction-Following Thai-English Machine Translation, AACL-IJCNLP 2026 Main Conference

Grounded is the supervised fine-tuning dataset built by Reference-Grounded Data Curation (RGDC) and used to train the ChindaMT family of instruction-following Thai-English translation models. Every constraint in a record was extracted from a reference translation that already satisfies it, so each target is feasible by construction.

• Records: 1,973,358 (1,172,061 en-th, 801,297 th-en), from 1,206,946 unique source pairs
• Format: Alpaca (instruction, input, output), in the exact form used for training
• License: CC-BY-SA 4.0

Fields

Field	Type	Description
instruction	string	translation directive, source text, and a Rules: block of one or more constraints
input	string	empty
output	string	translation that satisfies every listed constraint

Example

{
  "instruction": "Translate the given content into English.\nOriginal text:\nดังนี้ ฟ้า และ แผ่นดิน โลก และ บรรดา บริวาร ก็ ถูก สร้าง ขึ้น ให้ สำเร็จ\n### Response:\n\nRules:\n- Include the phrase 'vast array' in the translation.\n- Use a formal and archaic English tone.\n- Maintain the passive voice in the translation.",
  "input": "",
  "output": "Thus, the sky and the earth, the world and its vast array of attendants, were all brought to perfect completion."
}

Curation metadata

Per-record curation metadata (the extracted constraints, the verification questions, and the judge verdicts behind each row) is kept by the authors and can be published on request.

Provenance

Sources are ten public English-Thai parallel corpora, cherry-selected by Instruction-Following Difficulty in Phase 1 and constraint-augmented in Phase 2. Records per source corpus:

Corpus	Records
hplt	660,898
paracrawl	649,378
xlent	326,134
scb_2020	239,871
opensubtitles	63,048
bible	17,326
wikimedia	8,212
en-th-texts	8,035
tatoeba	392
elrc	64

Loading

from datasets import load_dataset

ds = load_dataset("iapp/ChindaMT-Grounded", split="train")

Citation

@misc{chayintr2026rgdc,
  title         = {Reference-Grounded Data Curation for Instruction-Following Thai-English Machine Translation},
  author        = {Chay-intr, Thodsaporn and Harnchang, Krittapad and Thabua, Mahannop and Viriyayudhakorn, Kobkrit and Theeramunkong, Thanaruk},
  year          = {2026},
  eprint        = {2609.34770},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.34770},
  note          = {Accepted at AACL-IJCNLP 2026 (Main Conference)}
}

License

CC-BY-SA 4.0, consistent with the ShareAlike terms of several source corpora. Credit the ChindaMT project and license derivatives under CC-BY-SA 4.0.