Tɛkyerɛma translator v3 (NLLB-600M)

Author: Prince Nasamu Alhassan

English to and from Twi, Ewe, Hausa, Dagbani and Kusaal. It continues tekyerema-nllb600m-v1 on all ten English-centred directions (18 Oct 2026). Kusaal is trained on a clean, hand-checked set; Twi and Hausa are capped at 120k pairs so they don't drown out the smaller languages, and Dagbani is repeated three times.

The rule for release was simple: the big gains on Dagbani and Kusaal could not cost more than 1 chrF++ on any FLORES direction for Twi, Ewe or Hausa. The worst drop was 0.73 (Ewe to English).

direction	test set	v1	v3
English → Twi	FLORES+	35.50	34.88
Twi → English	FLORES+	44.66	44.74
English → Ewe	FLORES+	39.43	39.60
Ewe → English	FLORES+	39.43	38.70
English → Hausa	FLORES+	52.10	53.83
Hausa → English	FLORES+	55.25	55.82
English → Dagbani	held out	32.58	37.52
Dagbani → English	held out	49.37	49.94
Kusaal → English	Kusaal Wikipedia	60.15	61.84
English → Kusaal	Kusaal Wikipedia	35.85	40.80
English → Kusaal, numbers	own set	26.92	35.00

chrF++, higher is better.

Use it

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
repo = "PrinceAlhassanNasamu/tekyerema-nllb600m-v3"
tok, model = AutoTokenizer.from_pretrained(repo), AutoModelForSeq2SeqLM.from_pretrained(repo).eval()
tok.src_lang = "eng_Latn"
enc = tok("Send fifty cedis to Kofi.", return_tensors="pt")
out = model.generate(**enc, forced_bos_token_id=tok.convert_tokens_to_ids("dag_Latn"), num_beams=2, max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True))

Language codes: twi_Latn, ewe_Latn, hau_Latn, dag_Latn, kus_Latn, eng_Latn.

Where it runs

The Tɛkyerɛma language service uses it for every direction except English → Twi, which stays on tekyerema-nllb600m-v2: v2 was tuned on the app's own replies ("data bundle", "payment", "contact") and still says them best.