---
language:
- en
- te
license: apache-2.0
task_categories:
- question-answering
- text-retrieval
- text-generation
pretty_name: Telugu QA Paraphrases
size_categories:
- 1K<n<10K
---

# Telugu QA Paraphrases

A synthetic multilingual query-rewriting dataset for evaluating retrieval robustness under Telugu-English code mixing.

## Dataset Description

This dataset extends an existing Telugu QA dataset by generating multiple query variants with increasing levels of Telugu-English code mixing.

Each example contains:

- `question` : Original English question
- `answer` : Ground-truth answer
- `level_0` : English paraphrase
- `level_1` : Light Telugu-English code mixing
- `level_2` : Moderate Telugu-English code mixing
- `level_3` : Heavy Telugu-English code mixing
- `level_4` : Romanized Telugu
- `level_5` : Telugu script

## Motivation

Multilingual embedding models are often evaluated on pure English or fully translated queries.

Real-world users frequently write:

- Telugu-English code mixed text
- Romanized Telugu
- Mixed-script queries

This dataset enables evaluation of retrieval robustness across progressively increasing code-mixing levels.

## Example

```json
{
  "question": "How do I apply for a passport in India?",
  "answer": "...",
  "level_0": "How can I apply for an Indian passport?",
  "level_1": "భారతదేశంలో పాస్పోర్ట్ ఎలా అప్లై చెయ్యాలి?",
  "level_2": "India-lo passport apply process enti?",
  "level_3": "భారత్ పాస్పోర్ట్ అప్లికేషన్ ఎలా చేయాలి?",
  "level_4": "India passport apply ela?",
  "level_5": "భారతదేశ పాస్పోర్ట్ కోసం ఎలా అప్లై చేయాలి?"
}
```

## Intended Uses

- Multilingual retrieval
- Dense retrieval evaluation
- RAG benchmarking
- Code-mixed query understanding
- Embedding model robustness analysis

## Source Dataset

This dataset was generated from:

https://huggingface.co/datasets/gurumurthy3/Legal-FAQ

Please cite and credit the original dataset creators.

## Generation

Query variants were generated using Sarvam-M and subsequently filtered for semantic preservation.

## Limitations

The generated paraphrases may contain:

- Imperfect transliteration
- Script mixing
- Translation artifacts
- Semantic drift in a small number of examples

Manual validation is recommended for benchmark creation.

## Citation

```bibtex
@dataset{telugu_qa_paraphrases,
  title={Telugu QA Paraphrases},
  author={Suhas Koheda},
  year={2026}
}
```