nlpctx/tamil-english-retrieval

Tamil-English Code-Mixed Retrieval Evaluation Dataset

This dataset is a publication-quality multilingual and code-mixed retrieval evaluation benchmark generated from parallel English-Tamil document datasets. It is designed to evaluate information retrieval systems, dense retrievers, and cross-lingual models on natural queries in various language formats.

Dataset Provenance & Source

• Original Source Dataset: mozhi-ai/tamil-corpus (Mozhi Tamil Corpus).
• English Translations: Machine-translated into English using AI4Bharat's IndicTrans2 translation model.
• Intermediate Corpus: nlpctx/tamil-english-corpus (Tamil-English Retrieval Corpus).
• Original Dataset License: Apache 2.0.
• Final Dataset License: Apache 2.0.

Dataset Statistics

• Total documents in original dataset (nlpctx/tamil-english-corpus): 4,591 documents.
• Documents processed from original dataset: 2,183 documents (representing the subset successfully completed by the pipeline).
• Total generated retrieval queries (final dataset rows): 9,299 query rows.
• Average query variants per document: 4.26 queries.

Dataset Summary

For each processed document from the Mozhi AI parallel corpus, realistic retrieval search queries (representing search intents for search engines, forums, or AI assistants) were generated in four variants:

• English (query_en): Standard English retrieval query.
• Tamil (query_ta): Native Tamil script retrieval query.
• Tamil Romanized (query_ta_romanized): Transliterated Latin-script query naturally typed.
• Tamil-English Code-Mixed (query_ta_codemix): Tanglish code-mixed query naturally mixing Tamil syntactic structure with English words.
• Answer Text (answer_text): Concise ground truth answer snippet extracted directly from the document.

Data Schema

Each row in the dataset contains the following fields:

Column	Type	Description
doc_id	string	Unique document identifier (corresponds to source corpus id)
query_en	string	English search query
query_ta	string	Native Tamil search query
query_ta_romanized	string	Transliterated Tamil search query (Tanglish)
query_ta_codemix	string	Code-mixed Tamil-English search query
answer_text	string	Snippet answering the query (ground-truth)
text	string	Full Tamil source document
english_doc	string	Full English parallel document
source	string	Source corpus identifier
url	string	URL metadata of source document
license	string	License of the source text
date_collected	string	Date when document was collected
language_score	float	Quality score of document language identification
quality_score	float	Document quality rating
domain	string	Domain classification of text
register	string	Document register style
title	string	Document title (if available)
tamil_char_ratio	float	Ratio of Tamil characters in native document
unique_word_ratio	float	Ratio of unique words in native document
avg_word_length	float	Average word length of native document
word_count	float	Word count of native document
char_count	float	Character count of native document
sentence_count	float	Sentence count of native document
pii_count	float	Detected PII count in document
has_code_switching	boolean	Whether source document contains code-switching
generation_model	string	LLM used to generate the retrieval queries
generation_timestamp	string	Timestamp of query generation

Generation & Quality Assurance

1. Dynamic Key Rotation & Model Fallback: Generation leveraged Gemma 4 models (gemma-4-26b-a4b-it, gemma-4-31b-it) and Gemini 3 models under rate-limit rotation.
2. Quality-preserving Pydantic Validators: Every generated row passed strict validations ensuring correct script blocks (Tamil Unicode vs. Latin script boundaries) and length checks.
3. Stopword-filtered Jaccard Similarity Deduplication: Rejects near-duplicate queries within and across document chunks using a Jaccard threshold of 0.85.

Citations

If you use this dataset, please cite the original datasets and models:

Mozhi Tamil Corpus

@misc{mozhi_ai_tamil_corpus_2026,
  title={mozhi-ai Tamil Corpus},
  author={mozhi-ai contributors},
  year={2026},
  url={https://huggingface.co/datasets/mozhi-ai/tamil-corpus}
}

IndicTrans2 Machine Translation Model

@article{gala2023indictrans,
  title={IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages},
  author={Jay Gala and Pranjal A. Chitale and A. K. Raghavan and Varun Gumma and Sumanth Doddapaneni and Aswanth Kumar M and Janki Nawale and Anupama Sujatha and Ratish Puduppully and Vivek Raghavan and Pratyush Kumar and Mitesh M. Khapra and Raj Dabre and Anoop Kunchukuttan},
  journal={Transactions on Machine Learning Research},
  year={2023}
}