ghanaopenai/gloss-mt-translategemma-4b

GLOSS-MT (TranslateGemma 4B + Glossary)

ghanaopenai/gloss-mt-translategemma-4b is an English → Asante Twi (Akan) translation model: a LoRA adapter for google/translategemma-4b-it to translate text in which dictionary terms have already been pretranslated in place. An offline English → Twi dictionary supplies the terms, so terminology can be changed by editing the dictionary instead of retraining the model.

• Author: Ghana Open AI
• Supported by: Ghana NLP
• Code: GhanaOpenAI/gloss-mt
• Live demo: Hugging Face Spaces: GLOSS-MT
• Training data: ghanaopenai/gloss-mt-parallel-en-twi
• Dictionary: 613,259 English → Twi phrases, one translation each (twi_dictionary_single.tsv.gz)

How it works

English text
  │  split into line-aligned chunks of at most 192 tokens
  ▼
spaCy noun phrases ──► offline dictionary lookup
  │
  ▼
found phrases replaced by their Twi translation in angle brackets, in place
  "Our people have courage, but courage without foundation is foolishness."
  "Our <nnipa> have <akokoduro>, but <akokoduro> without <fapem> is <nkwasiasɛm>."
  │
  ▼
TranslateGemma 4B + this LoRA adapter  en → ak (fine-tuned)  ──► Twi chunk with the terms in brackets
  │  brackets removed, chunks re-joined
  ▼
Twi text

Only phrases found in the dictionary are bracketed; the others stay plain English for the model to translate. The model learned to keep each bracketed term and translate the sentence around it, writing the term in brackets too, so the pipeline can strip them. Dot leaders of tables of contents are kept out of the model and restored unchanged. Editing the dictionary changes the terms the model writes, without retraining.

Benchmark (nsanku-mt-benchmark, English → Asante Twi)

Document-level chrF2 on the nsanku MT benchmark (84 documents, 4 domains; a language score is the mean of its per-dataset scores):

Model	chrF2	BLEU	Length ratio	Finance (3)	Folktales (75)	Watchtower (5)	UDHR (1)
Gemini 3.8 Flash	54.64	17.25	1.02	52.50	40.33	69.28	56.44
Google Translate	51.26	15.67	1.01	48.06	37.54	67.15	52.30
GLOSS-MT NLLB-200 3.3B	50.80	13.11	1.00	50.23	37.16	64.17	51.66
GLOSS-MT NLLB-200 1.3B	50.27	12.82	0.99	49.41	36.51	63.66	51.51
GLOSS-MT NLLB-200 600M	50.12	12.23	0.99	49.81	36.61	62.80	51.25
Khaya AI	49.91	12.79	1.04	47.74	36.22	65.78	49.92
GLOSS-MT TranslateGemma 4B	49.81	12.24	1.06	50.03	35.48	62.54	51.19
GLOSS-MT OPUS-MT en-tw	48.00	10.91	0.97	46.13	35.15	60.62	50.09

TranslateGemma has no Twi language code and, zero-shot, does not produce Twi; with this adapter (target code ak) and the dictionary it reaches 49.81 chrF2, its output running slightly long (length ratio 1.06).

Controllability

On 300 human-translated Asante Twi Bible passages (2,766 dictionary terms), the model writes 96.6% of the dictionary terms in its output (26.2% of the same terms appear without the dictionary). When every entry is replaced by a different term, 97.3% of the new terms appear and 1.9% of the old ones remain. With the dictionary it scores chrF2 37.15 on these passages, against 32.13 without it.

Training

• Data: 31,909 line-aligned chunks (≤ 192 tokens) of English–Twi documents. Each English noun phrase was aligned to its span in the gold Twi with Gemini, and the aligned phrases are bracketed on both sides (english_pretranslated → twi_bracketed), so the model learns to keep the bracketed terms.
• Method: LoRA r=32, α=64 (60M trainable parameters), bf16, AdamW, lr 2e-4, batch 8 × 4 gradient accumulation. The dictionary terms ride inside TranslateGemma's own translation prompt (source en, target ak).
• Checkpoint choice: every 1,000 steps the model translated 200 fixed validation chunks; the checkpoint with the best mean of BLEU, chrF2 and 100·e^(−loss) was kept: step 3,000 (~3 epochs; validation BLEU 33.73, chrF2 64.96). Decoding is greedy.

Usage

The model expects the pretranslated input that the GLOSS-MT pipeline builds:

pip install git+https://github.com/GhanaOpenAI/gloss-mt.git
python -m spacy download en_core_web_sm

from twi_translate.translategemma import TranslateGemmaTranslator

translator = TranslateGemmaTranslator("google/translategemma-4b-it",
                                      adapter="ghanaopenai/gloss-mt-translategemma-4b")
result = translator.translate("The court announced the decision yesterday.")
print(result["translation"])
print(result["lookup_results"])  # dictionary phrases used

The adapter loads with peft.PeftModel.from_pretrained on top of google/translategemma-4b-it (AutoModelForImageTextToText), with TranslateGemma's chat template (source_lang_code="en", target_lang_code="ak").

The GLOSS-MT models

GLOSS-MT adds an editable dictionary to pretrained translation models: dictionary terms are pretranslated in place in the source, and the model is fine-tuned once to use them.

Model	Base	Fine-tune	chrF2	Base chrF2
ghanaopenai/gloss-mt-nllb-3.3b	NLLB-200 3.3B	LoRA, merged	50.80	47.74
ghanaopenai/gloss-mt-nllb-1.3b	NLLB-200 1.3B	full	50.27	47.42
ghanaopenai/gloss-mt-nllb-600m	NLLB-200 distilled 600M	full	50.12	46.71
ghanaopenai/gloss-mt-translategemma-4b	TranslateGemma 4B	LoRA adapter	49.81	no Twi zero-shot
ghanaopenai/gloss-mt-opus-mt-en-tw	OPUS-MT en-tw	full	48.00	42.52

All of them can be tried in the GLOSS-MT Space.

Scope

Trained on paragraph- and document-level prose (news, reports, education, narratives). Isolated words or greetings without sentence context may translate poorly; give complete sentences.

License

The adapter is a derivative of TranslateGemma and is subject to the Gemma Terms of Use; the training data is CC-BY-NC-4.0, so use is non-commercial.