translate-windy-hplt-th-en — th to en

CTranslate2 INT8 bilingual translation model, derived from HPLT/translate-th-en-v2.0-hplt_opus (the HPLT project, CC-BY-4.0).

CC-BY-4.0 permits commercial use and modification and requires attribution; this notice provides it. Windstorm Labs did not train the base model.

Recovery note

HPLT publishes native Marian (.npz + .spm + .vocab) and marks its repos inference: false. An earlier conversion of this model to transformers format corrupted the weights, producing untranslatable output, and the model was written off. It was recovered on 2026-07-26 by converting the original .npz with the official Marian converter. The model was always fine; the conversion was not.

What was changed

A deliberately minimal LoRA fine-tune ("windy hallmark") merged into the base weights.

Rank / alpha	8 / 16
Learning rate / steps	2.5e-06 / 50
Target modules	q_proj, v_proj
Seed	42 (reproducible)
Training data	opus100_th-en (clinic cache), 2000 sentence pairs
Tensors modified	36
Max absolute weight delta	3.441e-05

Provenance

base   model.bin sha256  976dd1ada3cb37f0480eba2b6ec6038d6c2da5eca028055234def91cd42932a3
this   model.bin sha256  996e0b4d690210bb607a097e3ea96b24b5e20d7c48d1541583f867e45d33962c

Verified on the quantized artifact you download, not merely internal weights.

Evaluation

FLORES-200 devtest, 200 sentences, beam 4, CTranslate2 INT8 on CPU. Metrics: spBLEU (flores200 tokenizer) + chrF (word_order=0).

	spBLEU	chrF
this model	25.72	53.15
upstream, pre-fog	25.35	—

Limitations

• One language pair only (th to en).
• FLORES-200 is news and encyclopedic prose; it says little about conversational register.
• No human evaluation was performed.