Helax N3: A Sri Lankan First Large Language Model for AI Applications
Nimesh Harshana Balasooriya
NHB LK / Helax AI
Sri Lanka
2026
Abstract
Helax N3 is a Sri Lankan first large language model developed with the goal of improving natural language processing and generative AI capabilities for Sri Lankan users and applications. The model supports both Sinhala and English text generation and is distributed through the Hugging Face ecosystem under the MIT License.
The model is associated with the NVEagle/LocateAnything-Data dataset and is implemented within the NVIDIA NeMo ecosystem. Helax N3 is designed to investigate the development of language models that provide stronger support for Sri Lankan linguistic and cultural contexts while maintaining useful multilingual capabilities.
This paper presents the motivation, model characteristics, development methodology, dataset considerations, evaluation framework, limitations, and future research directions of Helax N3. Because a standardized public benchmark report for the current release has not yet been established, this paper does not claim unverified numerical performance results. Instead, it defines an evaluation framework that can be used to measure generation quality, instruction following, question answering, translation, and general language capabilities in future experiments.
The project demonstrates the potential of Sri Lankan-first language model development and provides a foundation for future research into Sri Lanka–centric artificial intelligence systems.
Keywords: Sri Lankan AI, Large Language Model, Helax AI, Helax N3, Natural Language Processing, Text Generation, Low-Resource Languages, Sinhala NLP
1. Introduction
Large language models have rapidly transformed natural language processing by enabling systems capable of text generation, question answering, summarization, translation, coding assistance, and conversational interaction.
However, the development and evaluation of large language models remains highly concentrated around languages with large digital datasets and extensive research resources. Sri Lanka, despite having rich linguistic diversity including Sinhala and Tamil alongside English usage, has comparatively fewer publicly available large-scale language resources, benchmarks, and specialized language models.
This creates an important research opportunity. A language model designed with a Sri Lankan-first perspective could provide better support for local users and applications while contributing to research on low-resource and underrepresented language ecosystems.
Helax N3 was developed as an experimental Sri Lankan-first language model intended to address this opportunity. The model supports Sinhala and English text generation and is distributed through Hugging Face.
The primary objectives of Helax N3 are:
- To investigate the development of a Sri Lankan-first language model.
- To provide a model capable of generating Sinhala and English text.
- To establish a foundation for Sri Lankan-specific AI applications.
- To encourage further research into Sri Lankan NLP.
- To provide an openly accessible model under the MIT License.
- To establish a foundation for future Sri Lankan language benchmarks and evaluation datasets.
2. Motivation
Sri Lanka presents several challenges and opportunities for modern NLP systems.
The country uses multiple languages in digital communication, including Sinhala, English, and Tamil, often in mixed or code-switched forms. Limited high-quality training resources and benchmarks make it difficult for general-purpose language models to consistently perform well in Sri Lankan contexts.
A model specifically developed for Sri Lankan linguistic and cultural contexts may therefore provide several advantages:
- Improved multilingual text generation for Sri Lanka.
- Better adaptation to local linguistic patterns.
- Greater accessibility for Sri Lankan developers.
- Support for conversational AI in local contexts.
- Support for educational applications.
- Support for local content generation.
- Increased research opportunities for Sri Lankan NLP.
Helax N3 is intended as a step toward this broader objective.
3. Model Overview
3.1 Model Name
Helax N3
3.2 Model Type
Large Language Model / Text Generation Model
3.3 Primary Languages
- Sinhala (
si)
- English (
en)
3.4 Framework
The model is associated with the NVIDIA NeMo ecosystem.
3.5 Primary Task
Text generation.
3.6 Dataset
The Hugging Face model metadata currently identifies:
NVEagle/LocateAnything-Data
as an associated dataset.
3.7 License
MIT License.
3.8 Model Repository
The public model repository is hosted on Hugging Face under:
nhblk123/helax-n3
4. Intended Use
Helax N3 is intended for research, experimentation, and development of AI applications involving Sri Lankan multilingual text.
Potential applications include:
- Sri Lankan conversational assistants.
- Question-answering systems.
- Educational AI tools.
- Content generation.
- Customer-support systems.
- Text summarization.
- Translation experiments.
- Sri Lankan NLP research.
- Developer experimentation.
- AI-powered information systems.
The model should be evaluated for a specific application before being deployed in production environments.
5. Data
The current model metadata identifies NVEagle/LocateAnything-Data as an associated dataset.
Dataset quality is particularly important for Sri Lankan-first language models because performance depends heavily on the quantity, diversity, accuracy, and linguistic quality of training data.
For future iterations of Helax N3, dataset development should consider:
- Sinhala linguistic diversity.
- Tamil language inclusion (future expansion).
- Formal and informal language usage.
- Sri Lankan terminology.
- Educational content.
- Government and public-service terminology.
- Technical terminology.
- Conversational language.
- Code-switching between Sinhala, English, and Tamil.
- Proper nouns and geographic entities.
- Cultural and contextual information.
The exact composition, size, filtering procedure, and preprocessing pipeline should be documented for each reproducible training release.
6. Methodology
6.1 Development Framework
Helax N3 is developed within the NVIDIA NeMo ecosystem.
NeMo provides infrastructure for developing, training, fine-tuning, and evaluating generative AI models.
6.2 Data Preparation
A robust training pipeline includes:
- Data collection from permitted sources
- Cleaning and normalization
- Language filtering
- Deduplication
- Tokenization optimized for multilingual input
7. Model Training
The training process follows:
Raw Data → Cleaning → Filtering → Tokenization → Training → Evaluation → Model Release
The model learns statistical language representations by predicting tokens based on context.
8. Sri Lankan First Design Considerations
Key design areas include:
8.1 Unicode Handling
Proper handling of Sinhala and multilingual scripts.
8.2 Tokenization
Efficient representation of Sinhala, English, and mixed text.
8.3 Code-Switching
Support for real-world Sri Lankan multilingual communication.
8.4 Local Context
Representation of Sri Lankan names, places, and cultural references.
9. Evaluation Methodology
9.1 Text Generation Quality
- Fluency
- Coherence
- Grammar
- Context consistency
9.2 Question Answering
- General knowledge
- Sri Lankan context
- Accuracy
9.3 Instruction Following
- Summarization
- Rewriting
- Classification
9.4 Translation
- Sinhala ↔ English
- Future: Tamil support
9.5 Human Evaluation
- Fluency
- Relevance
- Cultural appropriateness
- Accuracy
10. Benchmarks
No standardized public benchmark results are currently reported for Helax N3.
Future evaluations will compare Helax N3 with multilingual and Sri Lankan-relevant baselines using reproducible settings.
11. Results
No verified numerical benchmark results are included in this release.
Future results will include:
- Perplexity
- BLEU / ROUGE
- Exact Match
- F1
- Human evaluation scores
12. Limitations
- Limited Sri Lankan training resources
- Hallucination risk
- Lack of standardized benchmarks
- Multilingual complexity
- Need for improved dataset transparency
13. Ethical Considerations
Risks include:
- Misinformation
- Bias
- Cultural misrepresentation
- Privacy concerns
- Misuse in automated systems
14. Reproducibility
Future releases should document:
- Architecture
- Dataset details
- Training configuration
- Hardware
- Hyperparameters
- Evaluation scripts
15. Future Work
- Larger Sri Lankan multilingual datasets
- Sinhala + Tamil + English expansion
- Instruction tuning
- Human preference optimization
- Sri Lankan benchmark creation
- Multimodal AI capabilities
16. Impact on Sri Lankan AI
Helax N3 contributes to:
- Local AI development
- Research accessibility
- Educational tools
- Developer ecosystem growth
- Sri Lankan NLP advancement
17. Model Card Information
| Property | Value |
|---|
| Model | Helax N3 |
| Developer | NHB LK / Helax AI |
| Focus | Sri Lankan-first multilingual LLM |
| Languages | Sinhala, English |
| Framework | NVIDIA NeMo |
| Dataset | NVEagle/LocateAnything-Data |
| License | MIT |
| Repository | nhblk123/helax-n3 |
| Year | 2026 |
18. Conclusion
Helax N3 is a Sri Lankan-first large language model designed to support multilingual AI development for Sri Lankan applications. The model focuses on Sinhala and English generation while establishing a foundation for future expansion into broader Sri Lankan language support.
This release serves as a research foundation for Sri Lankan-centric AI systems rather than a final benchmarked model. Future work will focus on dataset expansion, evaluation standardization, and multilingual capability improvement.
19. References
(unchanged)
20. Citation
(unchanged)
21. Model Availability
Hugging Face: nhblk123/helax-n3
22. Acknowledgements
(unchanged)
If you want next step, I can also:
✅ convert this into NeurIPS / ACL paper format
✅ or make it look like a real Hugging Face model card (production ready)
✅ or add Tamil expansion + trilingual roadmap (very strong for Sri Lanka AI positioning)
