sheikhtijan/nllb-mandinka-english

NLLB-200 600M — Mandinka <-> English (v2)

Bidirectional Mandinka<->English translation, fine-tuned from facebook/nllb-200-distilled-600M using Bambara (bam_Latn) as a Mande-family proxy tag for Mandinka.

v2 results (test set)

Direction	BLEU
Mandinka -> English	22.8
English -> Mandinka	21.26

chrF++ by sentence length (MAN->EN): 1-15 words 37.2 · 16-30 words 45.5 · 31-50 words 42.7 · 51+ words 38.4. v2 was trained on a corpus expanded with 1,518 long-form pairs (avg 28 words, max 90); the 16-50 word range went from failure mode in v1 to the model's strongest.

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("sheikhtijan/nllb-mandinka-english")
model = AutoModelForSeq2SeqLM.from_pretrained("sheikhtijan/nllb-mandinka-english")

tok.src_lang = "bm_Latn"          # Mandinka input (Bambara proxy tag)
enc = tok("Suno man feeyaa.", return_tensors="pt")
out = model.generate(**enc,
                     forced_bos_token_id=tok.convert_tokens_to_ids("eng_Latn"),
                     max_length=192, num_beams=4)
print(tok.batch_decode(out, skip_special_tokens=True)[0])
# English -> Mandinka: set src_lang="eng_Latn" and force "bm_Latn" instead.

Training

7,315 parallel pairs trained bidirectionally (14,630 examples), 15 epochs, MAX_LENGTH 192, Adafactor + fp16 + gradient checkpointing, Kaggle T4.

Limitations

Strongest on 16-50 word sentences; short everyday/conversational vocabulary is the current weak spot. Orthography follows n-with-eng (ŋ) conventions. A JollofLab project, The Gambia.