🇸🇱 SierraLeoneNLP

Building open AI infrastructure for the languages of Sierra Leone.

SierraLeoneNLP is an open research and developer community focused on building, documenting, evaluating, and sharing Natural Language Processing (NLP), speech, and language AI resources for Sierra Leone.

Our goal is simple:

Make Sierra Leonean languages first-class languages in modern AI.

We believe researchers, developers, students, institutions, and communities should have access to the datasets, models, benchmarks, tools, and research needed to build useful AI systems for Sierra Leonean languages.

---

🌍 Mission

Sierra Leone is linguistically diverse, yet many of its languages remain significantly underrepresented in modern AI systems.

SierraLeoneNLP exists to help close this gap by creating an open ecosystem for research and development around Sierra Leonean languages.

We work to:

• 📚 Build high-quality language datasets
• 🗣️ Build speech datasets and speech technology
• 🤖 Develop and evaluate language models
• 🔊 Develop ASR and TTS resources
• 🌐 Build translation systems
• 📊 Create standardized benchmarks and evaluation datasets
• 🛠️ Develop open-source NLP tools and libraries
• 🔬 Support researchers experimenting with Sierra Leonean languages
• 👩🏾‍💻 Help developers build applications using these resources
• 🌍 Make Sierra Leonean language technology accessible to the global AI community

---

🎯 Vision

Our long-term vision is to establish a strong, open, and collaborative Sierra Leonean Language AI ecosystem.

We want a developer anywhere in the world to be able to visit SierraLeoneNLP and find:

• Datasets
• Pretrained models
• Evaluation benchmarks
• Documentation
• Research
• Training recipes
• APIs and tools
• Reproducible experiments
• Community contributions

for Sierra Leonean languages.

Ultimately, we want Sierra Leonean languages to be represented across the modern AI stack — from datasets and foundational models to speech assistants, translation systems, educational technology, search, accessibility tools, and conversational AI.

---

🇸🇱 Languages

SierraLeoneNLP aims to support the linguistic diversity of Sierra Leone.

Our work may include:

• Krio
• Mende
• Temne
• Limba
• Kono
• Loko
• Kissi
• Kuranko
• Susu
• Sherbro
• And other languages and varieties spoken in Sierra Leone.

The community will expand language coverage as more data, researchers, native speakers, and contributors become involved.

---

🧠 Areas of Research

Natural Language Processing

We develop resources for:

• Language modeling
• Text classification
• Named entity recognition
• Question answering
• Information extraction
• Text normalization
• Tokenization
• Linguistic analysis
• Conversational AI
• Instruction tuning
• Multilingual NLP

---

🗣️ Speech AI

We work toward better speech technology for Sierra Leonean languages.

Automatic Speech Recognition

Speech → Text

Projects may include:

• ASR datasets
• Transcription datasets
• Pretrained ASR models
• Evaluation benchmarks
• Speaker diversity research
• Speech data collection tools

Text-to-Speech

Text → Speech

Projects may include:

• TTS datasets
• Speaker datasets
• Pronunciation resources
• Pretrained TTS models
• Voice quality evaluation

Speech-to-Speech

Speech → Speech

We are interested in both modular and end-to-end speech systems that enable natural interaction with Sierra Leonean languages.

---

🌐 Machine Translation

We support research into translation between Sierra Leonean languages and other languages.

Examples include:

• English ↔ Krio
• English ↔ Mende
• English ↔ Temne
• Krio ↔ Mende
• Krio ↔ Temne
• Mende ↔ Temne
• And future multilingual combinations

Our goal is not simply to produce translation models.

We also want to build the high-quality parallel datasets, evaluation datasets, and benchmarks required to measure meaningful progress.

---

🤖 Language Models

SierraLeoneNLP supports research into language models capable of understanding and generating Sierra Leonean languages.

This includes:

• Pretrained language models
• Small language models
• Large language models
• Multilingual models
• Continued pretraining
• Instruction tuning
• Conversational datasets
• Language-specific evaluation
• Retrieval-augmented systems
• Language understanding
• Language generation

We encourage researchers to document both successful and unsuccessful experiments where possible so the community can learn collectively.

---

📊 Benchmarks & Evaluation

A major part of SierraLeoneNLP is building reliable benchmarks.

Having a model is not enough.

We need to know:

How well does it actually understand and generate Sierra Leonean languages?

We aim to develop standardized evaluation datasets and benchmarks for:

• ASR
• TTS
• Translation
• Language modeling
• Text understanding
• Question answering
• Instruction following
• Conversational AI
• Speech-to-speech systems

Evaluation may include both automated metrics and human evaluation by native speakers.

Potential metrics include:

• WER
• CER
• BLEU
• chrF
• COMET
• TER
• Perplexity
• Task-specific accuracy
• Human preference evaluation

Automated metrics will never be treated as the only measure of language quality.

---

📚 Datasets

High-quality data is one of the most important foundations of our work.

SierraLeoneNLP aims to publish and document datasets covering:

• Text
• Conversations
• Parallel translations
• Speech
• Transcriptions
• Pronunciation
• Linguistic annotations
• Instruction-following data
• Evaluation data

Every dataset should provide appropriate documentation describing:

• Data sources
• Collection methodology
• Preprocessing
• Licensing
• Intended use
• Limitations
• Known biases
• Speaker or contributor information where appropriate
• Ethical considerations

Data Access

Not every dataset will necessarily be open.

Some datasets may have restrictions because of:

• Consent
• Privacy
• Licensing
• Copyright
• Safety
• Cultural considerations
• Research agreements

Each dataset should clearly document its access conditions and license.

---

🔬 Research Principles

SierraLeoneNLP follows several principles.

1. Open by Default

When legally and ethically possible, we encourage open datasets, models, code, documentation, and research.

2. Quality Over Quantity

A large dataset is not automatically a good dataset.

We prioritize:

• Accurate data
• Diverse speakers
• High-quality transcriptions
• Consistent annotation
• Reproducibility
• Clear documentation

3. Native-Speaker Evaluation

Sierra Leonean language technology should ultimately be evaluated by people who actually understand and speak the languages.

4. Reproducibility

We encourage researchers to publish:

• Training configurations
• Preprocessing methods
• Evaluation procedures
• Model versions
• Dataset versions
• Limitations

where possible.

5. Responsible AI

We take privacy, consent, copyright, cultural context, and potential misuse seriously.

6. Community First

SierraLeoneNLP exists to serve the broader research and developer community rather than a single company or individual.

---

🤝 Contributions

SierraLeoneNLP welcomes contributions from:

• AI researchers
• NLP researchers
• ML engineers
• Software developers
• Linguists
• Native speakers
• Students
• Universities
• Independent researchers
• Open-source contributors
• Organizations working on African language technology

You do not need to be an expert to contribute.

You can contribute by:

• Submitting datasets
• Improving existing datasets
• Reporting errors
• Contributing code
• Training models
• Publishing benchmarks
• Improving documentation
• Evaluating models
• Providing native-speaker feedback
• Translating documentation
• Sharing research
• Creating tutorials
• Helping other contributors

---

🧑🏾‍💻 Community Standards

We aim to maintain a respectful and technically rigorous community.

Contributors should:

• Respect other contributors
• Respect language communities
• Document their work honestly
• Clearly identify limitations
• Avoid fabricated results
• Respect dataset and model licenses
• Protect private or sensitive information
• Properly attribute external work
• Avoid intentionally harmful uses of community resources

---

🔐 Responsible Open Source

SierraLeoneNLP is committed to open research, but openness must be balanced with:

• Privacy
• Consent
• Copyright
• Dataset licenses
• Safety
• Cultural considerations
• Responsible research practices

Therefore, projects may have different levels of openness.

A project may be:

• Fully open-source
• Open weights with restricted data
• Research-only
• Access-controlled
• Partially open
• Or unavailable for public release

Each project should clearly explain its licensing and access conditions.

---

🏗️ Project Structure

SierraLeoneNLP may host several types of repositories on Hugging Face.

Datasets

Examples:

SierraLeoneNLP/krio-text-corpus
SierraLeoneNLP/krio-english-parallel
SierraLeoneNLP/mende-text-corpus
SierraLeoneNLP/temne-speech