This model is a fine-tuned version of Helsinki-NLP/opus-mt-mul-en for Swahili-to-English translation, specifically optimized for child helpline call transcriptions in East Africa (Tanzania, Uganda, Kenya).

Developed by: BITZ IT Consulting Ltd
Project: OpenCHS (Open Child Helpline System)
Funded by: UNICEF Venture Fund
License: Apache 2.0

Training Data

The model was fine-tuned on a combination of:

NLLB Swahili-English parallel corpus (high quality, weight: 1.0)
CCAligned web-crawled parallel data (supplementary, weight: 0.7)
Total training samples: Approximately 50,000+ sentence pairs
Domain focus: Conversational Swahili from helpline contexts

Performance

Test Set (General Translation)

BLEU: 0.0195
chrF: 18.28
Improvement over baseline: +0.0%

Domain Evaluation (Call Transcriptions)

Domain BLEU: 0.0000
Domain chrF: 0.26
Domain COMET-QE: 0.0000

Domain metrics are evaluated on real 10-minute call transcriptions from child helplines.

Intended Use

Primary Use Case: Translating Swahili helpline call transcriptions to English for:

Case documentation and reporting
Quality assurance and supervision
Cross-border case referrals
Data analysis and insights

Languages: Swahili (source) → English (target)

Limitations:

Optimized for conversational Swahili (East African dialects)
May not perform well on formal/literary Swahili
Best for text lengths under 512 tokens
Requires post-editing for critical use cases

Usage

python

from transformers import MarianTokenizer, MarianMTModel

model_name = "YOUR_USERNAME/brendaogutu/sw-en-translation-test-3"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

# Translate
swahili_text = "Habari za asubuhi. Ninaitwa Amina na nina miaka 14."
inputs = tokenizer(swahili_text, return_tensors="pt", padding=True)
outputs = model.generate(**inputs, num_beams=5, max_length=256)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translation)

Training Details

Base Model: Helsinki-NLP/opus-mt-mul-en
Training Epochs: 0.1
Batch Size: 16 (effective: 64)
Learning Rate: 3e-05
Optimizer: AdamW
Hardware: NVIDIA GPU with FP16 mixed precision

Evaluation Methodology

Test Set: Random 5% split from training distribution
Domain Evaluation: Held-out set of real helpline call transcriptions (10-min calls)
Metrics: BLEU, chrF, COMET-QE (quality estimation)

Citation

bibtex

@software{openchs_translation_2025,
  author = {BITZ IT Consulting Ltd},
  title = {Swahili-English Translation Model for OpenCHS},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/YOUR_USERNAME/brendaogutu/sw-en-translation-test-3}
}

Contact

For questions or issues, please contact: brenda@openchs.org

This model is part of the OpenCHS project supporting child helpline services across East Africa.