Turaco-gem-mt-en-wes

Turaco-gem-mt-en-wes is a specialized translation model built on google/translategemma-4b-it and fine-tuned for high-quality machine translation from English to Cameroon Pidgin English (WES).

This model is part of the Turaco family, an initiative exploring modern translation-focused language models for low-resource and underrepresented African languages. By combining google/translategemma-4b-it’s multilingual capabilities with targeted fine-tuning on English–Cameroon Pidgin data, the model aims to produce translations that preserve meaning, tone, and natural expression.

Based on our internal evaluation, Turaco-gem-mt-en-wes is our best-performing Turaco model yet, representing an important step toward more capable and culturally relevant translation systems for Cameroon Pidgin.

Model details

Field	Value
Developed by	Clevaway
Model family	Turaco
Base model	google/translategemma-4b-it
Base revision	10042cb0e6e7fdce748996a71dc3dc432a4e0c89
Base parameter class	TranslateGemma 4B family; about 5B including vision components
Fine-tuning	BF16 LoRA supervised fine-tuning
LoRA rank / alpha	32 / 32
Direction	English (en) → Cameroon Pidgin (wes)
Model context configured	768 tokens
Maximum training sequence	512 tokens
Vision layers	Frozen and not evaluated for Cameroon Pidgin
Hardware	NVIDIA A100-SXM4-40GB (39.5 GB)

Cameroon Pidgin integration

TranslateGemma's native chat template maps language codes to human-readable language names. This checkpoint adds wes: Cameroon Pidgin to that mapping and otherwise preserves Google's translation prompt. The patched chat_template.jinja is included in this repo.

Training data

The model uses the already-cleaned fotiecodes/Turaco-NLLB-mt-en-wes-cleaned-dataset dataset at revision b95e28f77b36200d7be67a49513457d9c1f1696e. It does not rerun the original cleaning pipeline.

Split	Rows
Train	25,158
Validation	1,335
Test	1,354

Split source: published.

Training configuration

• epochs: 2
• learning rate: 0.0001
• micro-batch size: 8
• gradient accumulation: 4
• effective batch size: 32
• optimizer: adamw_torch_fused
• scheduler: cosine
• response-only loss: enabled
• deterministic seed: 42

Internal evaluation

These results are a same-corpus held-out diagnostic, not an external release benchmark.

Examples	chrF++	SacreBLEU	TER	Source copy	Empty output	Length ratio
1,354	43.09	25.54	71.18	0.00%	2.66%	0.977

Metric signatures are stored in internal_test_metrics.json. External TuracoBench v1 and speaker-reviewed results are pending. We do not describe this checkpoint as state of the art until those evaluations exist.

Usage

import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "fotiecodes/Turaco-gem-mt-en-wes"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [{
    "role": "user",
    "content": [{
        "type": "text",
        "source_lang_code": "en",
        "target_lang_code": "wes",
        "text": "What are you doing today?",
    }],
}]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_dict=True, return_tensors="pt"
).to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, do_sample=False, max_new_tokens=128)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Use Unsloth 4-bit loading on T4 and other float16-only GPUs.

Intended use

• reviewed English-to-Cameroon-Pidgin translation drafts;
• low-resource MT research and evaluation;
• localization prototyping;
• responsible commercial applications that independently satisfy the Gemma terms, applicable law, and the dataset's terms.

This model is not intended to be the sole translator for medical, legal, emergency, immigration, financial, or other high-stakes decisions.

Limitations

• The training corpus is narrow, noisy, and domain-skewed.
• Cameroon Pidgin has valid regional, dialect, and spelling variation.
• The model may copy English, omit content, add unsupported content, or overuse familiar training-domain phrasing.
• wes was not in TranslateGemma's original supported chat-template mapping.
• Image translation into Cameroon Pidgin was not trained or evaluated by this project.
• Internal metrics from the same source corpus can overestimate real-world quality.

License and notices

The model weights are governed by the Gemma Terms of Use. Commercial use is permitted subject to those terms and restrictions; this is not an unrestricted Apache or MIT model. See NOTICE, MODIFICATIONS.md, and the accompanying Gemma agreement copy.

This repository is a modified Gemma model derivative. Use and redistribution are subject to the Gemma Terms of Use and the Gemma Prohibited Use Policy.

Project code and model weights have different licenses. Dataset terms also apply separately.

Citation

@software{turaco_translategemma_2026,
  author = {fotiecodes},
  title = {Turaco-gem-mt-en-wes},
  year = {2026},
  url = {https://huggingface.co/fotiecodes/Turaco-gem-mt-en-wes}
}

@article{gemmatranslate2026,
  title = {TranslateGemma Technical Report},
  author = {Google Translate Research Team and others},
  year = {2026},
  url = {https://arxiv.org/abs/2601.09012}
}