Turaco-NLLB-mt-en-wes

Turaco-NLLB-mt-en-wes is a direction-specific checkpoint in the Turaco model family. It translates English → Cameroon Pidgin. It fine-tunes Meta's NLLB-200 distilled 600M encoder-decoder model with a dedicated wes_Latn language token and an extended SentencePiece vocabulary learned from the training split.

Model details

Item	Value
Base model	facebook/nllb-200-distilled-600M
Architecture	NLLB encoder-decoder; full supervised fine-tuning
Parameters after extension	615,657,472 (about 616M)
Direction	English → Cameroon Pidgin (eng_Latn → wes_Latn)
Language-token strategy	Added wes_Latn as a real special token; initialized from English and Tok Pisin embeddings
SentencePiece pieces added	571; trained on the training split only
Model-weight license	CC BY-NC 4.0, inherited from the NLLB base model

Use the tokenizer shipped with this checkpoint. Its vocabulary and language-token IDs differ from the original NLLB tokenizer. The saved generation configuration selects wes_Latn by default.

Training data and split

The model uses the cleaned MT560 dataset, derived from michsethowusu/english-cameroon-pidgin_sentence-pairs_mt560. The source data is CC BY 4.0. The released weights retain the base model's CC BY-NC 4.0 license. Cleaning normalizes Unicode and spacing, strips leading reference markers, filters empty and extreme-length pairs, and removes duplicate normalized pairs. Rows linked by the same normalized English or Cameroon Pidgin sentence stay in one split.

Raw pairs	Clean pairs	Train	Validation	Test
28,159	27,847	25,158	1,335	1,354

Training configuration

Item	Value
Requested / completed epochs	1 / 1.0
Optimizer steps	63
Peak learning rate	3e-05
Per-device batch / accumulation / effective batch	2 / 16 / 32
Maximum source / target tokens	160 / 192
Precision, optimizer	FP16, Adafactor
Gradient checkpointing, label smoothing	Yes, 0.1
Selection	Lowest validation loss; best checkpoint restored before export
Hardware	Tesla T4
Seed	42

Evaluation

These are internal same-corpus diagnostics, not an external benchmark. The grouped test partition comes from the same source corpus as training. External speaker-reviewed TuracoBench evaluation is pending. These figures do not establish state-of-the-art quality.

Test examples	chrF++	SacreBLEU	TER
300	29.32	9.38	80.65

Reproducible SacreBLEU signatures:

chrF++: nrefs:1|case:mixed|eff:yes|nc:6|nw:2|space:no|version:2.6.0
BLEU: nrefs:1|case:mixed|eff:yes|tok:13a|smooth:exp|version:2.6.0
TER: nrefs:1|case:lc|tok:tercom|norm:no|punct:yes|asian:no|version:2.6.0

The repository includes per-example predictions, metrics, training information and log, data audit, and split manifest.

Inference

import torch
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer

model_id = "fotiecodes/Turaco-NLLB-mt-en-wes"
tokenizer = NllbTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device).eval()
tokenizer.src_lang = "eng_Latn"
inputs = tokenizer('Please tell me when the bus leaves.', return_tensors="pt", truncation=True).to(device)
with torch.inference_mode():
    output = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids("wes_Latn"),
        max_new_tokens=192,
    )
print(tokenizer.decode(output[0], skip_special_tokens=True))

Intended use and limitations

Use for research and assisted translation between English and Cameroon Pidgin, with human review of important outputs. The training corpus is heavily religious-domain and contains noisy alignments. The model may copy English, omit or invent content, mishandle negation, and miss valid dialect and spelling variation. It has not been validated for unattended medical, legal, emergency, immigration, or financial use. Commercial use is restricted by the base model's noncommercial license.