nileagi/swahili-language-exposure

swahili-language-exposure

Dataset Summary

swahili-language-exposure is a large-scale Swahili (Kiswahili) corpus designed for language exposure and continued pretraining of language models.

Unlike instruction-tuning datasets, this dataset focuses on exposing models to natural Swahili usage across conversations, explanations, narratives, technical discussions, and mixed-domain text. The goal is to improve fluency, vocabulary coverage, syntax, and cultural grounding in Swahili.

This dataset is developed and maintained by NileAGI.

Dataset Purpose

This dataset is intended for:

• —Continued pretraining (DAPT / CPT)
• —Language exposure before instruction tuning
• —Improving Swahili fluency and coherence
• —Reducing English dominance in multilingual models
• —Low-resource language research

It is not optimized for instruction-following by default.

Dataset Format

The dataset is provided in JSON Lines (`.jsonl`) format.

Each line contains a single Swahili text sample, without enforced instruction–response structure:

{
  "text": "Maudhui ya asili kwa Kiswahili yaliyoandikwa au kuzungumzwa katika mazingira halisi."
}

Example

{"text":"Nilikuwa najifunza kuhusu mitandao ya neva na jinsi inavyotumika katika utambuzi wa picha, lakini changamoto kubwa ilikuwa kupata data iliyosawazika."}

Supported Tasks

• —Language modeling
• —Text generation
• —Conversational fluency improvement
• —Vocabulary and grammar learning
• —Domain adaptation for Swahili
• —Foundation training for downstream fine-tuning

Language

• —Swahili (Kiswahili) — primary language (sw)
• —Natural code-switching with English technical terms may appear
• —Informal and semi-formal registers are both present

Data Sources

The dataset was curated from:

• —Educational explanations and mentoring sessions
• —Informal dialogues and narrative text
• —Mixed-domain Swahili content reflecting real usage

All data has been anonymized and cleaned.

Data Preprocessing

The following preprocessing steps were applied:

• —Filtering of extremely short or low-quality text
• —Normalization of whitespace and encoding
• —Preservation of natural language flow and code-switching

No artificial instruction templates were added.

Intended Use

Primary Use

• —Continued pretraining (CPT/DAPT) for Swahili
• —Language exposure before instruction tuning
• —Improving multilingual model balance

Secondary Use

• —Linguistic analysis
• —Swahili NLP benchmarking
• —Data augmentation for low-resource research

Out-of-Scope Uses

This dataset is not intended for:

• —Direct instruction fine-tuning (see companion instruct dataset)
• —Preference learning or RLHF
• —Surveillance, profiling, or harmful content generation

Biases and Limitations

• —Informal conversational Swahili is more common than formal prose
• —Technical domains may be overrepresented
• —Regional phrasing reflects contributor backgrounds

Users should consider complementary corpora for broader coverage.

Ethical Considerations

• —Personally identifiable information has been removed or masked
• —No sensitive personal attributes are intentionally included
• —Released strictly for responsible AI research

License

This dataset is released under the MIT License.

Citation

If you use this dataset, please cite:

@dataset{swahili_language_exposure,
  title = {swahili-language-exposure: A Swahili Language Exposure Dataset},
  author = {NileAGI},
  year = {2026},
  publisher = {Hugging Face}
}