CoolFace
Modelpublic

openchs/lg-opus-mt-multi-en-synthetic-v1

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face
1likes12downloads
Model Card

opus-mt-lg-en-finetuned - Version 1

Model Description

Fine-tuned Luganda-English translation model optimized for child helpline services. Trained on 42,510 child-welfare conversations with crisis detection and emotion preservation. Achieves 0.616 BLEU score (422% improvement over baseline). Specialized for counseling, family support, and crisis intervention contexts with cultural sensitivity for Ugandan child communication patterns. Includes safety filters and harmful content detection. 8 epochs training on Helsinki-NLP/opus-mt-mul-en base model with custom tokens for crisis/emotional contexts.

Model Name: opus-mt-lg-en-finetuned Version: 8 Task: Translation Framework: pytorch Export Method: transformers MLflow Experiment: translation-lg-en Training Date: 2025-08-02 05:14:25

Model Details

Architecture

  • —Model Type: Not specified
  • —Base Model: Helsinki-NLP/opus-mt-mul-en
  • —Framework: pytorch
  • —Language(s): lg,en
  • —License: apache-2.0
  • —Fine-tuned from: Helsinki-NLP/opus-mt-mul-en

Intended Uses & Limitations

Intended Uses

Child helpline real-time translation, crisis intervention support, counseling sessions, family support communications, educational guidance for Ugandan children and adolescents

Limitations

Best on conversational/emotional content, may struggle with technical terms. 512 token limit. Enhanced emotional sensitivity may over-interpret neutral text. Requires <2s response time for crisis situations.

Training Data

Dataset

  • —Name: luganda-parallel.jsonl
  • —Size: 50012
  • —Languages: lg,en

Training Configuration

Hyperparameters

ParameterValue
Team Config NotesIncreased Batch by 2, epochs by 4 and max length * 2
Total Parameters77518848
Trainable Parameters76994560
Vocab Size64172
Max Length512
Language Pairlg-en
Language NameLuganda
Model NameHelsinki-NLP/opus-mt-mul-en
Dataset Config Primary Datasetcustom
Dataset Config Custom Datasets['dataset/custom/luganda/converteddata.jsonl', 'dataset/custom/luganda/lugandaparallel.jsonl']
Dataset Config Validation Split0.15
Dataset Config Max SamplesNone
Training Config Learning Rate4e-05
Training Config Batch Size4
Training Config Num Epochs12
Training Config Max Length256
Training Config Weight Decay0.01
Training Config Warmup Steps3000
Evaluation Config Metrics['bleu', 'chrf', 'meteor']
Evaluation Config Test Size150
Team Config Assigned DeveloperMarlon
Team Config Prioritymedium

Performance Metrics

Evaluation Results

MetricValue
Train Samples108613.0000
Validation Samples19167.0000
Total Samples127780.0000
Baseline Bleu0.1683
Baseline Chrf37.6363
Loss0.0793
Grad Norm4.5997
Learning Rate0.0000
Epoch11.0000
Eval Loss0.2365
Eval Bleu0.5028
Eval Chrf69.4387
Eval Runtime2114.2653
Eval Samples Per Second9.0660
Eval Steps Per Second2.2670
Bleu Improvement0.3345
Bleu Improvement Percent198.8019
Chrf Improvement31.8024
Chrf Improvement Percent84.4992
Train Runtime51057.2887
Train Samples Per Second25.5270
Train Steps Per Second6.3820
Total Flos47003705065930752.0000
Train Loss0.1499
Final Eval Loss0.2365
Final Eval Bleu0.5028
Final Eval Chrf69.4387
Final Eval Runtime2114.2653
Final Eval Samples Per Second9.0660
Final Eval Steps Per Second2.2670
Final Epoch11.0000

Environmental Impact

  • —Hardware Type: NVIDIA GPU with 4GB+ VRAM
  • —Training Duration: 7 hours

Usage

Installation

bash
pip install transformers torch

Example - Translation

python
from transformers import pipeline

# Load the model
translator = pipeline("translation", model="openchs/lg-opus-mt-multi-en-synthetic-v1")

# Translate text
result = translator("Nsobola okuba n’ennamba yo ey’essimu??")
print(result[0]["translation_text"])

MLflow Tracking

  • —Run ID: 212ff762127b4d569dcf850b23f6268b
  • —Experiment ID: 22
  • —Experiment Name: translation-lg-en
  • —Tracking URI: http://192.168.10.6:5000

Model Card Authors

BITZ AI TEAM

Model Card Contact

info@bitz-itc.com

Citation

If you use this model, please cite:

bibtex
@misc{opus_mt_lg_en_finetuned_8,
  title={opus-mt-lg-en-finetuned v8},
  author={BITZ AI TEAM},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/marlonbino/opus-mt-lg-en-finetuned}
}

This model card was automatically generated from MLflow metadata and user inputs.