- Lozi โ English Neural Translation Models
- 1. Introduction
- 2. Model Overview
- 3. Model Architecture
- 4. Training Environment
- 5. Training Data
- 6. Training Results
- 7. Translation Examples
- 8. Model Evaluation
- 9. Installation and Usage
- 10. Intended Applications
- 11. Known Limitations
- 12. Responsible AI and Ethical Considerations
- 13. Deployment Considerations
- 14. Development Roadmap
- 15. Licence and Usage Terms
- 16. Citation
- 17. Acknowledgments
- 18. Version History
- 19. Contact and Support
Lozi โ English Neural Translation Models
Advancing Local-Language AI for Digital and Agricultural Inclusion in Zambia
Developed by: Netagrow Technologies Limited
Model Version: 1.0
Training Date: 17 August 2026
Release Date: 17 August 2026
Reporting Period: August 2026
Development Platform: Kaggle
Model Status: Research and Evaluation
Licence: Proprietary โ All Rights Reserved
1. Introduction
Netagrow Technologies Limited has developed bidirectional neural machine translation models for Lozi (Silozi) and English, supporting research into local-language artificial intelligence, digital inclusion, and language preservation in Zambia.
Lozi is a Bantu language widely spoken in Zambia's Western Province and neighbouring parts of Southern Africa. It is an important language of communication, education, and cultural expression within the Barotseland region.
Despite its significance, Lozi remains comparatively underrepresented in many modern language technology systems.
This project seeks to address that gap through the development of two translation models:
- English โ Lozi: Translating English text into Lozi.
- Lozi โ English: Translating Lozi text into English.
Both models build on Meta AI's NLLB-200-distilled-600M architecture and were fine-tuned using parallel LoziโEnglish language datasets.
The models also form part of Netagrow's broader research into making agricultural knowledge, climate information, and digital advisory services accessible through local languages.
Project Objectives
- Develop neural translation capabilities for English and Lozi.
- Improve digital access to information for Lozi-speaking communities.
- Support research and development in low-resource African languages.
- Explore multilingual agricultural advisory applications.
- Contribute to language preservation through practical AI technologies.
2. Model Overview
| Attribute | English โ Lozi | Lozi โ English |
|---|---|---|
| Model type | Neural machine translation | Neural machine translation |
| Base architecture | NLLB-200-distilled-600M | NLLB-200-distilled-600M |
| Parallel sentence pairs | 5,429 | 5,515 |
| Training examples | 4,886 | 4,963 |
| Test examples | 543 | 552 |
| Training epochs | 15 | 15 |
| Training steps | 4,590 | 4,665 |
| Initial training loss | 8.359 | 5.274 |
| Final training loss | 1.126 | 1.929 |
| Loss reduction | 87% | 63% |
| Training duration | 42m 56s | 48m 10s |
Key Characteristics
- Bidirectional translation using separate directional checkpoints.
- Transformer-based encoder-decoder architecture.
- Fine-tuning of a pretrained multilingual model.
- Support for short and medium-length text.
- FP16 mixed-precision training.
- Focus on conversational and culturally relevant language.
- Potential integration into multilingual agricultural technologies.
Evaluation note: The reported loss reductions demonstrate learning during training. They do not independently establish translation accuracy.
3. Model Architecture
Base Model
The translation models were fine-tuned from:
facebook/nllb-200-distilled-600M
NLLB stands for No Language Left Behind, a multilingual machine translation research initiative developed by Meta AI.
The distilled model provides a pretrained multilingual foundation that can be adapted to lower-resource translation tasks.
Technical Specifications
| Component | Specification |
|---|---|
| Architecture | Sequence-to-sequence Transformer |
| Approximate parameters | 600 million |
| Encoder layers | 12 |
| Decoder layers | 12 |
| Hidden dimension | 1,024 |
| Attention heads | 16 |
| Feed-forward dimension | 4,096 |
| Tokenizer | SentencePiece-based tokenizer |
| Fine-tuning method | Full model fine-tuning |
| Maximum configured input length | 128 tokens |
| Training precision | FP16 mixed precision |
The models were not trained from scratch. They were initialised from a pretrained checkpoint and adapted to LoziโEnglish translation.
This approach reduces the training data and computing resources required compared with developing an equivalent model entirely from scratch.
Language Configuration
The training documentation identifies the following language codes:
- English:
eng_Latn - Lozi:
loz_Latn
The uploaded tokenizer and checkpoint configuration should be checked to confirm how these identifiers are represented and whether target-language generation requires additional configuration.
4. Training Environment
The models were trained using GPU resources provided through Kaggle.
Hardware
| Resource | Configuration |
|---|---|
| Platform | Kaggle Notebooks |
| GPU | NVIDIA Tesla P100 |
| GPU memory | 16 GB class |
| Operating system | Linux |
| Compute approach | GPU-accelerated fine-tuning |
Software
| Component | Version / Library |
|---|---|
| Python | 3.12.12 |
| PyTorch | 2.8.0 |
| CUDA | 12.6 |
| Hugging Face Transformers | 4.x |
| Tokenisation | SentencePiece |
| Supporting libraries | datasets, accelerate, evaluate |
Training Configuration
base_model: facebook/nllb-200-distilled-600M
training:
epochs: 15
per_device_batch_size: 4
gradient_accumulation_steps: 4
effective_batch_size: 16
learning_rate: 0.00003
weight_decay: 0.01
warmup_steps: 500
max_sequence_length: 128
mixed_precision: fp16
evaluation:
strategy: final_checkpoint
train_test_split: 90/10
random_seed: 42
Training Duration
| Model | Duration |
|---|---|
| English โ Lozi | 42 minutes 56 seconds |
| Lozi โ English | 48 minutes 10 seconds |
| Combined model training | 91 minutes 6 seconds |
The original project records also describe approximately 96 minutes of total training-related activity. The difference may include additional processing overhead and should be confirmed against the training logs.
5. Training Data
Dataset Overview
The models were fine-tuned using parallel text containing corresponding English and Lozi expressions.
The dataset was developed from language-learning and linguistic materials, including dictionaries, conversational phrases, and cultural expressions.
Dataset Composition
The project documentation describes the following approximate source categories:
| Content category | Reported proportion |
|---|---|
| Dictionaries and lexicographic material | 60% |
| Conversational phrases | 25% |
| Cultural expressions and proverbs | 15% |
These proportions are drawn from the project documentation and should be verified against the final dataset manifest.
Dataset Statistics
| Statistic | English โ Lozi | Lozi โ English |
|---|---|---|
| Total sentence pairs | 5,429 | 5,515 |
| Training examples | 4,886 | 4,963 |
| Test examples | 543 | 552 |
| Source CSV size | 382.9 KB | 462.2 KB |
| Approximate source vocabulary | 3,800 English words | 3,200 Lozi words |
Data Preparation
Reported preprocessing activities included:
- UTF-8 encoding standardisation.
- Whitespace normalisation.
- Punctuation standardisation.
- Removal of incomplete translation pairs.
- Duplicate and near-duplicate filtering.
- Review of selected conversational expressions.
- Inclusion of culturally relevant phrases and proverbs.
Dataset Limitations
The datasets are relatively small compared with large-scale commercial machine translation corpora.
Potential gaps include:
- Regional Lozi dialects.
- Contemporary vocabulary.
- Agricultural terminology.
- Technical and scientific expressions.
- Code-switching between English and Lozi.
- Longer conversational contexts.
A complete source register, permissions record, and independently reviewed evaluation dataset remain important future documentation priorities.
6. Training Results
English โ Lozi
| Metric | Result |
|---|---|
| Initial training loss | 8.359 |
| Final training loss | 1.126 |
| Approximate loss reduction | 87% |
| Training steps | 4,590 |
| Epochs | 15 |
| Training duration | 42m 56s |
The training process showed a substantial reduction in loss, indicating that the model learned patterns from the parallel training data.
Lozi โ English
| Metric | Result |
|---|---|
| Initial training loss | 5.274 |
| Final training loss | 1.929 |
| Approximate loss reduction | 63% |
| Training steps | 4,665 |
| Epochs | 15 |
| Training duration | 48m 10s |
The reverse-direction model also demonstrated convergence, although the reported training loss remained higher.
Interpretation
These results support the conclusion that both fine-tuning runs made optimisation progress.
They do not establish that one model is more accurate than the other, because training loss alone is not a directly comparable translation-quality benchmark.
Independent evaluation is required before making reliable claims about translation accuracy or production readiness.
7. Translation Examples
The following examples were recorded in the original development documentation.
English โ Lozi
| English input | Reported Lozi output |
|---|---|
| Good morning | Muzuhile |
| How are you? | Mu cwani? |
| I am fine | Ni iketile |
| Thank you | Luitumezi |
| Where are you going? | Uya kai? |
These examples were reported as acceptable during preliminary review. A reproducible native-speaker evaluation protocol has not yet been published.
Lozi โ English
| Lozi input | Reported English output | Observation |
|---|---|---|
| Ni pilile | I am fine | Reported as acceptable |
| U pila cwani? | How do you live? | Overly literal |
| Muzibubile | Open air | Reported mistranslation |
| Ni lumela | I believe | Context-dependent |
| U ya kwanaa? | Where are you going? | Reported as acceptable |
Observed Translation Challenges
The preliminary examples suggest several areas requiring further improvement:
- Idiomatic translation.
- Homophone and polysemy handling.
- Cultural context.
- Grammatical variation.
- Ambiguous expressions.
8. Model Evaluation
Available Metrics
| Evaluation measure | Status |
|---|---|
| Training loss | Available |
| Loss reduction | Available |
| Qualitative examples | Available |
| BLEU | Not reported |
| chrF | Not reported |
| METEOR | Not reported |
| Independent human evaluation | Not formally reported |
| Dialect-specific evaluation | Not formally reported |
| Agricultural-domain benchmark | Not yet reported |
Future Evaluation Framework
Netagrow intends to introduce a more comprehensive evaluation process covering:
Automated Evaluation
- BLEU.
- chrF.
- Additional suitable translation metrics.
- Held-out test-set evaluation.
Human Evaluation
- Translation adequacy.
- Fluency.
- Cultural appropriateness.
- Terminology accuracy.
- Context preservation.
Domain Evaluation
- Agricultural instructions.
- Weather advisories.
- Pest and disease terminology.
- Seasonal farming guidance.
- Farmer questions and responses.
The objective is to assess real-world translation quality rather than relying exclusively on training metrics.
9. Installation and Usage
Requirements
Install the necessary Python libraries:
pip install "transformers>=4.40,<5" torch sentencepiece
English โ Lozi Translation
import torch
from transformers import (
AutoTokenizer,
AutoModelForSeq2SeqLM
)
CHECKPOINT = "./model_english_to_lozi"
tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT)
model = AutoModelForSeq2SeqLM.from_pretrained(CHECKPOINT)
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)
model.eval()
text = "Good morning, how are you?"
inputs = tokenizer(
text,
return_tensors="pt",
truncation=True,
max_length=128
).to(device)
with torch.inference_mode():
outputs = model.generate(
**inputs,
max_new_tokens=128,
num_beams=4
)
translation = tokenizer.decode(
outputs[0],
skip_special_tokens=True
)
print("Translation:", translation)
Lozi โ English Translation
Load the corresponding reverse-direction checkpoint:
CHECKPOINT = "./model_lozi_to_english"
tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT)
model = AutoModelForSeq2SeqLM.from_pretrained(CHECKPOINT)
The remaining inference steps are the same, but the input must be in Lozi.
Batch Translation
sentences = [
"Good morning",
"Thank you",
"Where are you going?"
]
inputs = tokenizer(
sentences,
return_tensors="pt",
padding=True,
truncation=True,
max_length=128
).to(device)
with torch.inference_mode():
outputs = model.generate(
**inputs,
max_new_tokens=128,
num_beams=4
)
translations = tokenizer.batch_decode(
outputs,
skip_special_tokens=True
)
for source, translated in zip(sentences, translations):
print(f"{source} -> {translated}")
Important: The examples assume the appropriate directional checkpoints are available in the specified local directories. Verify the uploaded repository structure and tokenizer language configuration before using Hugging Face Hub paths directly.
10. Intended Applications
General Language Technology
The models can support research and prototyping in:
- Multilingual applications.
- Educational language-learning tools.
- Digital content localisation.
- Linguistic research.
- Conversational interfaces.
- Cultural and language documentation.
Agricultural Technology
Netagrow is exploring the application of local-language AI within its agricultural technology ecosystem.
Netagrow App
Potential applications include:
- Translation of crop-production guidance.
- Livestock-management information.
- Multilingual agricultural interfaces.
- Translation assistance for farmer questions.
- Local-language digital advisory experiences.
Neta Alert
Potential applications include:
- Local-language seasonal agricultural information.
- Weather and climate advisory translation.
- Multilingual risk communication.
- Agricultural early-warning messages.
- WhatsApp-based agricultural guidance.
These represent intended or potential applications. They do not establish that these particular checkpoints are already deployed in production.
11. Known Limitations
Literal Translation
The Lozi โ English model may generate literal translations instead of natural English equivalents.
Contextual Ambiguity
Words and phrases with multiple meanings may be incorrectly interpreted when insufficient context is provided.
Regional Dialects
The dataset primarily represents standard Lozi. Regional varieties have not been comprehensively evaluated.
Code-Switching
Mixed LoziโEnglish input may produce inconsistent results.
Technical Vocabulary
Specialised terminology may not be adequately represented in the training data.
Input Length
The training configuration used a maximum sequence length of 128 tokens. Longer passages should be segmented before translation.
Cultural Expressions
Proverbs, idioms, and culturally specific expressions may require human interpretation.
Translation Reliability
The models have not yet undergone comprehensive independent benchmarking and should not be assumed to provide consistently accurate translations.
12. Responsible AI and Ethical Considerations
Language Inclusion
The project aims to improve the representation of Lozi in digital language technologies.
Cultural Sensitivity
Translations involving cultural expressions should be reviewed by knowledgeable Lozi speakers.
Fairness
Formal demographic and dialect-specific fairness evaluation has not yet been documented.
Future evaluation should examine whether performance varies across dialects, contexts, and language registers.
Privacy
The models were not reported as being trained using differential privacy.
Consequently, the possibility of training-data memorisation cannot be ruled out without dedicated testing.
Production applications should implement suitable privacy and data-handling safeguards.
Human Oversight
Human review is strongly recommended for consequential translations.
The models should not be relied upon without qualified review for:
- Medical instructions.
- Legal documents.
- Financial commitments.
- Emergency communications.
- Safety-critical agricultural advice.
- Pesticide handling instructions.
13. Deployment Considerations
Model Size
| Item | Reported size |
|---|---|
| English โ Lozi model | Approximately 2.46 GB |
| Lozi โ English model | Approximately 2.46 GB |
| Combined uncompressed weights | Approximately 4.92 GB |
| Compressed model archive | Approximately 2.18 GB per model |
Hardware Considerations
GPU acceleration is recommended for faster inference.
Actual memory requirements and inference performance depend on:
- Hardware configuration.
- Batch size.
- Precision.
- Sequence length.
- Decoding strategy.
- Software environment.
Optimisation Opportunities
Future work may investigate:
- INT8 quantisation.
- Reduced-precision inference.
- ONNX export.
- Inference batching.
- Model distillation.
- Resource-efficient deployment.
No independent deployment benchmark is presented in this release.
14. Development Roadmap
- Expand parallel sentence datasets.
- Improve Lozi โ English idiomatic translation.
- Add contextual disambiguation examples.
- Strengthen agricultural terminology coverage.
- Include regional dialect variations.
- Improve code-switching handling.
- Conduct formal native-speaker evaluation.
- Publish BLEU and chrF results.
- Document dataset provenance and permissions.
- Benchmark inference performance.
- Explore quantisation for lower-resource deployment.
- Evaluate integration into Netagrow agricultural services.
15. Licence and Usage Terms
Copyright ยฉ 2026 Netagrow Technologies Limited. All Rights Reserved.
The project documentation identifies the fine-tuned models and associated materials as proprietary.
Commercial Use
Commercial use requires appropriate written authorisation from Netagrow Technologies Limited, subject to applicable upstream licence terms.
Redistribution
Redistribution of Netagrow-owned model materials requires authorisation under the applicable licence.
Research and Educational Use
Non-commercial research, educational evaluation, and compatibility testing may be permitted under the applicable access terms.
Upstream Licensing
The models are derived from Meta's NLLB architecture. Any distribution or licensing arrangements must remain compatible with the upstream model licence and the permissions associated with the training datasets.
For licensing or collaboration enquiries, contact Netagrow Technologies Limited.
Website: https://www.netagrow.com
16. Citation
If you use these models in academic research or publications, please cite:
@misc{netagrow_lozi_2026,
title = {Lozi--English Bidirectional Neural Machine Translation Models},
author = {{Netagrow Technologies Limited}},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/netagrow/en-lozi-transformer},
note = {Fine-tuned NLLB-based translation models for Lozi and English}
}
17. Acknowledgments
Netagrow acknowledges:
- Meta AI for the NLLB research and pretrained model architecture.
- Kaggle for providing GPU computing resources.
- Hugging Face for model-hosting and machine-learning tools.
- The Lozi-language knowledge represented in the source materials.
These acknowledgments do not imply formal organisational endorsement.
18. Version History
| Date | Version | Description |
|---|---|---|
| 17 August 2026 | 1.0 | English โ Lozi model training completed |
| 17 August 2026 | 1.0 | Lozi โ English model training completed |
| August 2026 | 1.0 | Model documentation and project reporting |
Version 1.0 Highlights
- Bidirectional translation development.
- Fine-tuning of pretrained multilingual Transformer models.
- 5,429 English โ Lozi parallel sentence pairs.
- 5,515 Lozi โ English parallel sentence pairs.
- 15 training epochs per direction.
- Final training losses of 1.126 and 1.929.
- Preliminary qualitative translation examples.
- Documented limitations and future evaluation priorities.
19. Contact and Support
Organisation: Netagrow Technologies Limited
Country: Zambia
Website: https://www.netagrow.com
Hugging Face: https://huggingface.co/netagrow
Model Repository: https://huggingface.co/netagrow/en-lozi-transformer
Model Version: 1.0
Release Date: 17 August 2026
Status: Research and Evaluation
Developed by Netagrow Technologies Limited โ advancing local-language AI, agricultural innovation, and digital inclusion in Zambia.