---
pretty_name: "Swahili Small Clean Corpus (Demo)"
language:
  - sw
license: "cc-by-sa-4.0"
task_categories:
  - text-generation
  - fill-mask
language_creators:
  - found
multilinguality: monolingual
size_categories:
  - 50K<n<100K
dataset_size: "≈52.8 MB"
tags:
  - swahili
  - wikipedia
  - text
  - cleaned
  - deduplicated
---

# Swahili Small Clean Corpus (Demo)

**One-line summary**  
A small high-quality Swahili text corpus (≈58k documents, ~52.8 MB) extracted from Swahili Wikipedia and public news sources, cleaned, deduplicated, and split into `train` / `validation` / `test`. Suitable for fine-tuning and small-scale experiments.

## Dataset composition
- **Languages:** Swahili (`sw`)
- **Total rows:** ≈58,386
- **Size:** ≈52.8 MB (parquet)
- **Splits:** train (≈46.7k rows), validation (≈5.84k rows), test (≈5.84k rows)
- **Fields:**  
  - `title`  article  
  - `text` — cleaned article text (UTF-8)  
  - `source` — source identifier ( `swahili_wikipedia`, `bbc_swahili`)  
  - `url` — URL

## How the dataset was created
1. Downloaded Swahili Wikipedia dump and scraped selected public Swahili news pages.  
2. Extracted text from XML/HTML; cleaned wiki markup and HTML using regex-based cleaning.  
3. Performed language filtering (kept Swahili), basic quality filtering (min length, special-char heuristics), and exact deduplication.  
4. Sampled and saved top-quality documents; created train/validation/test splits.  
5. Saved in Hugging Face `datasets` format (Parquet/Arrow) and uploaded to the Hub.

## Provenance & license
- Primary source: Swahili Wikipedia (CC BY-SA). If any other sources were included (news sites), please check their terms for redistribution.  
- **License:** CC-BY-SA-4.0 (confirm this is compatible with all included sources; if you included non-free sources, change this field accordingly).

## Limitations & safety
- The dataset may contain bias inherent to source texts. Evaluate for harmful content before deploying.  
- PII redaction was performed with simple regex heuristics; manual audit is recommended for production use.

## Example: Loading the dataset
```python
from datasets import load_dataset
ds = load_dataset("marcoharuni95/swahili-text-corpus")
print(ds)
print(ds['train'][0]['text'][:400])

Example: Minimal fine-tuning snippet (illustrative only)

This is a short example showing how you might begin fine-tuning a causal model (adapt to your environment; use appropriate compute and licenses):

# Example only — adapt model, tokenizer and training loop to your infra.
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

ds = load_dataset("maroharuni95/swahili-text-corpus")
train_ds = ds['train']
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", trust_remote_code=True)
# tokenization & training steps omitted for brevity — see transformers + PEFT docs



Citation

If you use this dataset in research, please cite it as:

@misc{swahili_small_demo_2025,
  title = {Swahili Small Clean Corpus (Demo)},
  author = {Your Name},
  year = {2025},
  howpublished = {Hugging Face Dataset},
  note = {https://huggingface.co/datasets/YOUR_USERNAME/YOUR_REPO_NAME}
}


Contact

If you have questions or find issues, contact: marcoharuni95@gmail.com.com.