ChindaMT-CoreEval

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

800 sample English-Thai evaluation set across 5 deployment-relevant domains (government, finance, technology, UI, miscellaneous). Primary evaluation suite for the ChindaMT model family.

• plain split: 400 plain-translation samples (200 en-to-th + 200 th-to-en).
• constrained split: 400 separate samples (200 en-to-th + 200 th-to-en) with a Rules: block of 1–4 rules drawn from 5 constraint categories (content, numerical, stylistic, format, linguistic).

Each row carries the source text, the formatted instruction, and pre-generated predictions from ChindaMT-4B, ChindaMT-2B, and ChindaMT-0.8B. No reference translation is shipped; the suite is designed for pairwise LLM-judge or human evaluation.

Dataset structure

Each row in plain.jsonl and constrained.jsonl has the following fields:

Field	Type	Description
id	string	Unique row identifier
direction	string	en->th or th->en
domain	string	One of gov, finance, tech, ui, misc
context_type	string	sentence or paragraph
source_text	string	The raw source text
instruction	string	Formatted prompt fed to the model
phenomena_tags	list[string]	Optional content tags (e.g., has_latin)
rules	list[string]	Constraint rules (constrained split only)
verifiable_checks	list[object]	Optional machine-checkable rule checks, each with a type and a value (constrained split only)
split	string	plain or constrained
prediction_chindamt_4b	string	Output from ChindaMT-4B
prediction_chindamt_2b	string	Output from ChindaMT-2B
prediction_chindamt_08b	string	Output from ChindaMT-0.8B

Loading

from datasets import load_dataset

plain = load_dataset("iapp/ChindaMT-CoreEval", "plain", split="train")
constrained = load_dataset("iapp/ChindaMT-CoreEval", "constrained", split="train")

Provenance

CoreEval is sampled from a held-out split of the RGDC source pool, ten public English-Thai parallel corpora (bible, paracrawl, elrc, hplt, en-th-texts, opensubtitles, scb_2020, tatoeba, wikimedia, xlent), length-stratified and assigned to five deployment-relevant domains during curation. The split was held out before Phase 1. The source corpus of individual rows was not recorded.

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

Released under CC-BY-SA 4.0, consistent with the ShareAlike terms of several source corpora. When using this dataset, please credit the ChindaMT project and license any derivatives under CC-BY-SA 4.0.

iApp AI Research