Kord Translate MBART50 (EN↔TH V2)

This model is a fine-tuned version of facebook/mbart-large-50-many-to-many-mmt for Thai–English bidirectional translation. It was developed as part of the rationale-free distillation study described in the paper "Teaching the Student to Skip the Homework: Rationale-Free Distillation for Thai-English Translation".

Despite being trained on only ~8,000 distilled pairs, this model achieves competitive performance with purpose-built Thai–English systems (Typhoon Translate, ChindaMT) on BLEU and COMET metrics, particularly in the Thai→English direction.

Model Details

Summary

Description

mBART-50 is a multilingual sequence-to-sequence model pretrained on 50 languages. This variant is fully fine-tuned on approximately 8,000 rationale-free distilled translation pairs generated by DeepSeek-V4-Flash, covering both English→Thai and Thai→English directions.

"Rationale-free" means that while the teacher model (DeepSeek-V4-Flash) was prompted to reason through a multi-step translation procedure (identifying word senses, classifying register, selecting honorifics, and rewriting naturally), only the final translation was retained for student training. The student never sees the teacher's reasoning trace—only the final output.

Intended Uses & Limitations

Intended Use

This model is intended for Thai–English machine translation in both directions. It is suitable for:

• General-purpose Thai↔English translation
• Low-resource translation research and benchmarking
• Applications where a lightweight, efficient seq2seq model is preferred

Limitations

• Training data size: The model was fine-tuned on only ~8,000 pairs, which is small relative to typical MT training sets.
• Domain coverage: Training data draws from Wikipedia, books, exams, legal texts, IT content, and forum data; performance may degrade on highly specialized or colloquial domains.
• Comparison to larger models: While competitive with purpose-built Thai–English systems on BLEU and COMET, this model trails on chrF++ and BERTScore in the en→th direction.

How to Use

With Transformers Pipeline

from transformers import pipeline

translator = pipeline(
    "translation",
    model="KordAI/Kord-Translate-MBART50-ENTH-V2",
    tokenizer="KordAI/Kord-Translate-MBART50-ENTH-V2",
    src_lang="en_XX",
    tgt_lang="th_TH"
)

# English → Thai
result = translator("Ring also settled a lawsuit with competing security company, the ADT Corporation.")
print(result[0]['translation_text'])

With AutoModelForSeq2SeqLM (Explicit Language Codes)

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained("KordAI/Kord-Translate-MBART50-ENTH-V2")
tokenizer = AutoTokenizer.from_pretrained("KordAI/Kord-Translate-MBART50-ENTH-V2")

# English → Thai
src_text = "Ring also settled a lawsuit with competing security company, the ADT Corporation."
inputs = tokenizer(src_text, return_tensors="pt", src_lang="en_XX")
forced_bos_token_id = tokenizer.lang_code_to_id["th_TH"]
outputs = model.generate(**inputs, forced_bos_token_id=forced_bos_token_id)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# Thai → English
src_text = "ริงยังได้ดำเนินการระงับข้อพิพาทกับบริษัทรักษาความปลอดภัย ADT Corporation ซึ่งเป็นคู่แข่งอีกด้วย"
inputs = tokenizer(src_text, return_tensors="pt", src_lang="th_TH")
forced_bos_token_id = tokenizer.lang_code_to_id["en_XX"]
outputs = model.generate(**inputs, forced_bos_token_id=forced_bos_token_id)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

The model was fine-tuned on a rationale-free distillation dataset of approximately 8,000 bidirectional pairs (~4,000 per direction). Source text was drawn from nine corpora spanning formal and informal registers:

Teacher Model

Translations were generated by DeepSeek-V4-Flash with extended thinking enabled. The teacher was prompted to follow a four-stage reasoning procedure (literal meaning identification, genre/formality classification, vocabulary selection, natural rewriting) but to output only the final translation—omitting all reasoning traces.

Training Procedure

Evaluation

Metrics

The model was evaluated on the FLORES devtest set (1,012 samples per direction) using:

• BLEU – n-gram overlap
• chrF / chrF++ – character-level F-score
• BERTScore – semantic similarity via BERT embeddings
• COMET – learned neural MT evaluation metric

Results

Key Findings

• Dramatic improvement over the untuned mBART-50 baseline (en→th BLEU: 0.58 → 1.46; th→en BLEU: 9.02 → 17.30) — nearly tripling BLEU in both directions.
• Competitive performance with purpose-built Thai–English systems (Typhoon Translate, ChindaMT) on BLEU and COMET metrics, particularly in the th→en direction, despite using only ~8,000 training pairs.
• The largest relative gains from rationale-free distillation occur for the weakest base model, suggesting that a small distilled sample primarily corrects for missing domain exposure rather than adding new reasoning capability.

Sample Translations

English → Thai

Thai → English

Citation

If you use this model, please cite the following paper:

@article{jangjit2026teaching,
  title={Teaching the Student to Skip the Homework: Rationale-Free Distillation for Thai-English Translation},
  author={Jangjit, Naphon and Komsang, Jeerawat and Boran, Kord C.},
  journal={Zenodo},
  year={2026},
  doi={10.5281/zenodo.21325610},
  url={https://doi.org/10.5281/zenodo.21325610}
}

Model Card Authors

Naphon Jangjit, Jeerawat Komsang, Kord C. Boran (KordAI)

Additional Information

Related Models

Acknowledgements

This work was supported by KordAI. We thank the developers of DeepSeek-V4, mBART-50, and the FLORES benchmark for making their resources available.