victorachede/tiv-translator

TRANSLTR — English → Tiv Translation Model

Fine-tuned for Tiv, a Benue-Congo language spoken by ~4 million people in Benue State, Nigeria.
Built by Victor Achede under Black Sheep Co.

---

Model Summary

Property	Detail
Base model	Helsinki-NLP/opus-mt-en-mul
Task	Machine translation (EN → TIV)
Language pair	English → Tiv (tiv)
Architecture	MarianMT (transformer seq2seq)
Training data	Custom curated EN↔TIV parallel corpus (Bible-domain, conversational)
Fine-tuning epochs	10
Batch size	32
Hardware	NVIDIA T4 (Google Colab)
Framework	HuggingFace Transformers 4.x

---

Why This Exists

Tiv is one of Nigeria's major languages — spoken by millions across Benue State and the diaspora — yet it has zero representation in any major NLP benchmark, translation API, or pretrained multilingual model.

Google Translate doesn't support it. DeepL doesn't support it. NLLB-200 doesn't support it.

This model is the first step toward changing that. It is part of TRANSLTR, a real-time spoken language translation system being built to bridge Tiv speakers into the digital world — starting with live event translation at the GCK Benue conference, July 2025, IBB Square, Makurdi.

---

Usage

from transformers import MarianMTModel, MarianTokenizer

model_id  = "victorachede/tiv-translator"
tokenizer = MarianTokenizer.from_pretrained(model_id)
model     = MarianMTModel.from_pretrained(model_id)

def translate(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True)
    out    = model.generate(**inputs, max_length=128, num_beams=4)
    return tokenizer.decode(out[0], skip_special_tokens=True)

print(translate("For God so loved the world."))
print(translate("The Lord is my shepherd."))
print(translate("Ask and it shall be given unto you."))

---

Training Data

The dataset is a custom-built parallel corpus of English–Tiv sentence pairs, assembled specifically for this project. Sources include:

• Bible text (primary domain) — Luke, John, Acts (priority books)
• Conversational Tiv — everyday phrases and common expressions
• Manual curation by native Tiv speakers

Dataset size at time of training: 28,987 verified pairs
Dataset is actively growing. Model will be retrained as corpus expands.

---

Limitations

• Low-resource reality: 233 pairs is a starting point, not a ceiling. Outputs improve meaningfully with each dataset expansion.
• Domain: currently strongest on Bible-register English. Colloquial or technical text may produce weaker results.
• Tokenizer mismatch: the base MarianMT tokenizer was not trained on Tiv — subword segmentation of Tiv tokens is imperfect at this stage. A Tiv-native tokenizer is on the roadmap.
• This is v0.1. It is not production-ready. It is a proof-of-concept that this problem is solvable.

---

Roadmap

• Fine-tune MarianMT base on EN→TIV (28,987 pairs)
• Expand dataset to 100,000+ pairs
• Train Tiv-native SentencePiece tokenizer
• ElevenLabs Tiv voice clone integration (TTS output)
• Groq Whisper STT pipeline (speech input)
• Live demo at GCK Benue, June 2025
• Mobile SDK for offline Tiv translation

---

Citation

If you use this model or dataset in your research, please cite:

@misc{achede2025tivtranslator,
  author    = {Victor Achede},
  title     = {TRANSLTR: English-Tiv Neural Machine Translation},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/victorachede/tiv-translator}
}

---

About

Built by Victor Achede, founder of Black Sheep Co. — a technology holding company based in Makurdi, Benue State, Nigeria.
TRANSLTR is one of several products under active development targeting African language infrastructure, live event technology, and low-resource NLP.

"If it doesn't exist, build it."