Designed for pretraining and SFT of Marathi-aware Small Language Models that should understand and generate the way Marathi is commonly written online — Roman-script Marathi naturally mixed with English terms.

Example

Question:
  Yashwant Dev kon hote exactly — sangeetkar, kavi, ki donhi?

Answer:
  Yashwant Dev he Marathi music madhe ek legendary figure hote. Te primarily
  sangeetkar (music composer) mhanun famous zale, but tyanchya creative range
  madhe kavita writing pan included hote ...

Dataset Overview

The dataset contains approximately 1.09 million synthetic QA pairs covering a broad range of topics represented in Marathi Wikipedia.

Each source article is used to generate multiple questions spanning different dimensions of understanding, including:

Factual knowledge
Causal reasoning
Processes and explanations
Significance and impact
Comparisons
Contemporary context
Human and biographical perspectives
Common misconceptions
Contextual understanding

The dataset is intended to help Marathi language models learn code-mixed Romanized Marathi, rather than treating Marathi as exclusively Devanagari text.

Source

Base corpus: wikimedia/wikipedia, 20231101.mr — approximately 94K Marathi Wikipedia articles
Source language: Marathi
Writing style: Romanized Marathi with natural English code-mixing
Questions per article: 12 diverse question types
Answer length: approximately 300–500 words

The source articles and their corresponding Wikipedia URLs are retained in the dataset metadata for provenance.

Schema

Column	Type	Description
`id`	string	Unique pair ID in the format `row{NNNNNN}-q{NNN}`
`row_idx`	int32	Source Wikipedia article index
`source_title`	string	Wikipedia article title in Devanagari
`source_url`	string	Source Wikipedia article URL
`question`	string	Code-mixed Roman Marathi question
`answer`	string	Code-mixed Roman Marathi answer
`est_tokens`	int32	Estimated token count of the answer

Quick Load

from datasets import load_dataset

ds = load_dataset(
    "atx-labs/marathi-codemix-wiki-qa-1.4M",
    split="train"
)

print(ds[0]["question"])
print(ds[0]["answer"])

Intended Use

This dataset is intended for:

Marathi language model pretraining
Supervised fine-tuning (SFT)
Code-mixed language modeling
Romanized Marathi generation
Marathi QA systems
Research on Indic code-mixing and transliteration
Building Marathi-aware conversational models

Limitations

This is a synthetic dataset generated from Wikipedia-derived source material. As a result:

Generated answers may contain factual errors or omissions.
The dataset inherits potential biases and gaps present in the underlying Wikipedia corpus.
Code-mixed Marathi reflects one style of Romanized Marathi and should not be considered a standardized writing system.
The dataset should be reviewed or filtered further for high-stakes applications.
Synthetic QA pairs should not be treated as independently verified factual sources.

Licence & Attribution

The underlying Wikipedia source material is licensed under CC BY-SA 4.0.

The synthetic Question–Answer dataset is released under CC BY-SA 4.0, with attribution and ShareAlike requirements applying to the dataset and adaptations of it.

Please review the applicable terms and attribution requirements of the upstream Wikipedia content when redistributing or creating derivative works.

Contributors

Guruprasad Pathak — @grpathak22

About OpenXFI

This dataset is released as part of OpenXFI, an open-source Indic AI initiative by ATX Labs.

OpenXFI builds and openly releases foundational models, benchmarks, and datasets for Indian languages — with the goal of making high-quality Indic AI accessible to researchers, developers, and institutions across the country.

Explore the wider release, including our Indic evaluation benchmark and open datasets, at openxfi.com.

Citation

If you use this dataset in your research or applications, please cite the dataset and the OpenXFI initiative:

@misc{marathiminglishqa2026,
  title        = {Marathi Minglish QA: Synthetic Code-Mixed Marathi Question--Answer Dataset},
  author       = {Pathak, Guruprasad},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/atx-labs/marathi-codemix-wiki-qa-1.4M}},
  note         = {Released as part of OpenXFI by ATX Labs}
}

@misc{openxfi2026,
  title        = {OpenXFI: Shaping India's AI, Together},
  author       = {{ATX Labs}},
  year         = {2026},
  url          = {https://openxfi.com}
}

Contact

Website: openxfi.com
ATX Labs: atxlabs.ai
Issues & feedback: Open a discussion on this dataset's Community tab.

Downloads last month: 6

Total file size:

1.63 GB