SynhalaAI/sinhala-song-lyrics

---
language:
- si
license: mit
size_categories:
- 1K<n<10K
task_categories:
- text-generation
- mask-generation
dataset_info:
features:
- name: title
dtype: string
- name: lyrics
dtype: string
- name: artist
dtype: string
- name: source
dtype: string
---

# 🎵 SynhalaAI — Ultimate Sinhala Song Lyrics Dataset (Gold Mix)


## Dataset Description

The **SynhalaAI Lyrics Corpus** is a meticulously engineered, high-fidelity dataset of Sinhala song lyrics. It was designed specifically to train Large Language Models (LLMs) and advanced tokenizers on the poetic, colloquial, and structured linguistic patterns of the Sinhala language.

Unlike standard web-scraped datasets that are littered with English guitar chords, metadata, and HTML noise, this dataset has undergone a **"Sinhala-First" Surgical Scrub**. It is pristine, purely Sinhala, and spans decades of Sri Lankan music history.

### 🌟 Key Features
* **AI-Grade Cleanliness:** All English characters, guitar chords (e.g., `[Am]`, `G`, `Dm`), beat markers (e.g., `Beat: 4/4`), invisible Unicode ghost characters, and web links have been surgically removed.
* **Strict Deduplication:** Songs are aggressively deduplicated using a 150-character lyric fingerprinting algorithm combined with title cross-referencing.
* **Generational Span:** Contains everything from classic 1960s hits to fresh 2024-2026 releases.

## 🏗️ The 3-Pronged Data Strategy

This Master Dataset is a conglomerate of the best Sinhala archives available, gathered through a three-phase data engineering pipeline:

1. **The Mega Conglomerate (History Archive):** Aggregated data from 8+ of the largest open-source Sinhala search engine projects on GitHub, bringing in the bulk of classic and popular songs.
* *Sources:* Repositories by DilanSachi, BhanukaD610, Rashmini, Tharindu KW, rpjayasekara, Nirmal Gamage, and Thilina TLM.
2. **Linguistic & Metaphor Integration:** Integrated high-quality, human-curated data from Kaggle (`pasinduudawatta/sinhala-songs-and-metaphors`).
3. **DIY Fresh Mixer (2024-2026):** A custom, polite web scraper targeted at the "Gold Standard" (`Sinhalasongbook.com`) to extract the absolute latest releases omitted from older static archives.

## 📊 Dataset Structure

The dataset is provided as a unified JSON array (`sinhala_lyrics_MASTER.json`) and a flat text corpus (`sinhala_lyrics_corpus_MASTER.txt`) optimized for tokenizer training (BPE/SentencePiece).

### Data Fields
* `title` *(string)*: The title of the song (usually in Sinhala or Singlish).
* `lyrics` *(string)*: The deeply cleaned, Sinhala-only text of the song lyrics.
* `artist` *(string)*: The artist or singer of the song.
* `source` *(string)*: The original source of the data (e.g., `github_archive`, `kaggle_metaphors`, `fresh_scraper`).

### Sample Entry
```json
{
"title": "Bambareku Anduwa",
"lyrics": "බඹරෙකු ඇඬුවා මලක් ළඟදි බඹරෙකු ඇඬුවා\nඇඬුවා... මලක් ළඟදි බඹරෙකු ඇඬුවා \nඇහේ කඳුලු ඇති ඒ බඹරිඳු දැක\nඅඬන්නෙ ඇයිදෝ මල ඇසුවා... මල ඇසුවා...",
"artist": "Niranjala Sarojini",
"source": "fresh_scraper"
}