README.md

metadata

language:
  - ta
license: cc-by-sa-4.0
task_categories:
  - text-generation
  - fill-mask
tags:
  - tamil
  - corpus
  - nlp
  - llm-training
  - indic
  - dravidian
pretty_name: mozhi-ai Tamil Corpus
size_categories:
  - 1K<n<10K

mozhi-ai Tamil Corpus

A continuously-updated, high-quality Tamil language corpus for LLM training. Collects text from classical literature, Wikipedia, news, and web sources.

Corpus Stats (updated 2026-05-29)

Metric	Value
Total Documents	8,682
Total Characters	4,959,482
Total Words	553,873
Deduplicated	0 removed
Quality Filtered	0 removed
Last Updated	2026-05-29

Sources Breakdown

Source	Documents
hf:ai4bharat/IndicCorpV2	8,155
wikipedia	488
project_madurai	39

Licenses

License	Documents
CC-0	8,155
CC-BY-SA-4.0	488
public-domain	39

Domains

Domain	Documents
news	8,155
encyclopedia	488
classical_literature	39

Registers

Register	Documents
formal	8,643
classical	39

Data Fields

Field	Type	Description
id	string	Unique document ID (SHA-256 content hash)
text	string	The Tamil text content
source	string	Data source identifier
url	string	Source URL (when available)
license	string	License of the source data
date_collected	string	Collection date (ISO 8601)
domain	string	Content domain: news, encyclopedia, classical_literature, web
register	string	Language register: classical, formal, colloquial
title	string	Document title (Wikipedia, Madurai)
language_score	float	fastText Tamil confidence (0-1)
quality_score	float	Quality heuristic score (0-1)
tamil_char_ratio	float	Ratio of Tamil script chars to total non-whitespace
unique_word_ratio	float	Ratio of unique words to total words
avg_word_length	float	Average characters per word
word_count	int	Total word count
char_count	int	Total character count
sentence_count	int	Sentence count (based on punctuation)
pii_count	int	Detected PII instances (emails, phones, IPs)
has_code_switching	bool	Contains significant English (Tanglish)

Processing Pipeline

Every document passes through:

1. Language Detection — fastText lid.176.bin (threshold: 0.7)
2. Unicode Normalization — NFKC + Tamil-specific normalization (indic_nlp_library)
3. Exact Deduplication — SHA-256 content hashing
4. Quality Filtering — length, Tamil character ratio, uniqueness, boilerplate detection
5. Annotation — domain, register, text stats, PII detection, code-switching

Sources

• IndicCorpV2 (AI4Bharat) — Tamil news and articles (CC-0)
• allenai/c4 — Web-crawled Tamil text (ODC-BY)
• Tamil Wikipedia — Encyclopedia articles (CC-BY-SA-4.0)
• Project Madurai — Classical Tamil literature (Public Domain)

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("mozhi-ai/tamil-corpus")

# Stream without downloading
ds = load_dataset("mozhi-ai/tamil-corpus", streaming=True)

# Access the text
for example in ds["train"]:
    print(example["text"][:100])

Citation

If you use this corpus, please cite:

@misc{mozhi-ai-tamil-corpus,
  title={mozhi-ai Tamil Corpus},
  author={mozhi-ai contributors},
  year={2026},
  url={https://huggingface.co/datasets/mozhi-ai/tamil-corpus}
}

Contributing

This is an open project. Contributions welcome at GitHub.

---

Built with mozhi-ai — automated Tamil corpus pipeline