Dataset Card for Chitrakshara: A Large Multilingual Multimodal Dataset for Indian languages

Dataset Description

We introduce the Chitrakshara dataset series, covering 11 Indian languages sourced from Common Crawl. It comprises:

(1) Chitrakshara-IL (Interleaved), a large-scale interleaved image-text web documents
(2) Chitrakshara-Cap, image captioning dataset based on corresponding alt-text descriptions

Supported Languages

• Hindi
• Bengali
• Tamil
• Malayalam
• Telugu
• Marathi
• Kannada
• Gujarati
• Punjabi
• Oriya
• Assamese

Data Fields

We follow the same format as OBELICS in our dataset for uniformity. An example of a sample looks as follows:

{
    'images': [
        None,
        'https://ananttvlive.com/static/c1e/client/82638/uploaded/bd327ae26fd0ec3bed3b6ed250c343d1.jpg?width=789&height=605&resizemode=4',
        None
    ],
    'metadata': '[null, { "alt_text": "sdf", "document_url": "https://ananttvlive.com/Country/the-mela-committee-has-created-a-food-zone-at-the-garba/cid12523684.htm", "format": "jpeg", "formatted_filename": "bd ae fd ec bed b ed c d", "original_height": 605, "original_width": 789, "rendered_height": 153, "rendered_width": 200, "src": "https://ananttvlive.com/static/c1e/client/82638/uploaded/bd327ae26fd0ec3bed3b6ed250c343d1.jpg?width=789&height=605&resizemode=4", "unformatted_src": "https://ananttvlive.com/static/c1e/client/82638/uploaded/bd327ae26fd0ec3bed3b6ed250c343d1.jpg?width=789&height=605&resizemode=4" }, null]',
    'general_metadata': '{"url": "https://ananttvlive.com/Country/the-mela-committee-has-created-a-food-zone-at-the-garba/cid12523684.html", "warc_filename": "crawl-data/CC-MAIN-2023-50/segments/1700679100800.25/warc/CC-MAIN-20231209040008-20231209070008-00281.warc.gz", "warc_record_offset": 120398051, "warc_record_length": 17512}',
    'texts': [
        'मेला समिति ने इस बार गरबा स्थल पर फूड जोन बनाया है। यहां बड़ी संख्या में युवाओं ने चटपटे व्यंजनों का लुत्फ उठाया।',
        None,
        'गुजराती गरबा धुन तारा बिना श्याम मने एकालडू चलो के साथ युवाओं के हुजूम ने गुजराती लय में गरबा किया तो सिंधी गीत मुंहिंजो खटी आयो खैर सां होजमालो की धुन पर सिंधी स्टाइल में डांडिया करते हुए युवाओं को हुजूम झूम उठा।...मेला समिति के अध्यक्ष मनीष दरियानी के अनुसार शनिवार को गायक कलाकार शुभम नाथानी गरबे में शामिल होंगे।'
    ]
}

Similar to OBELICS, each sample is composed of the same 4 fields: images, texts, metadata, and general_metadata. images and texts are two lists of the same size, where for each index, one element and only one is not None. For example, for the interleaved web document <image_1>text<image_2>, we would find [image_1, None, image_2] in images and [None, text, None] in texts.

The images are replaced by their URLs, and the users need to download the images. This repo could be a good starting point.

metadata is the string representation of a list containing information about each of the images. It has the same length as texts and images and logs for each image relevant information such as original source document, unformatted source, alternative text if present, etc.

general_metadata is the string representation of a dictionary containing the URL of the document, and information regarding the extraction from Common Crawl snapshots.

Author Statement

We acknowledge that Chitrakshara may reflect inherent biases present in online content, as the dataset is sourced from the internet. Nonetheless, its multilingual and inclusive composition marks a significant step toward enhancing the accessibility of diverse languages, cultures, and communities for training India-centric Vision-Language Models (VLMs). While we have taken rigorous measures to ensure the accuracy and legality of the dataset, we cannot guarantee its absolute completeness or correctness. Consequently, the authors assume no liability for any potential legal or ethical concerns, including but not limited to copyright infringement, privacy violations, or the misuse of sensitive information.

Terms of Use

By using the dataset, you agree to comply with the original licenses of the source content as well as the dataset license. Additionally, if you use this dataset to train a Machine Learning model, you agree to disclose your use of the dataset when releasing the model or an ML application using the model.

Citation

@inproceedings{khan2025chitrakshara,
  title={Chitrakshara: A Large Multilingual Multimodal Dataset for Indian languages},
  author={Khan, Shaharukh and Faraz, Ali and Ravi, Abhinav and Nauman, Mohd and Sarfraz, Mohd and Patidar, Akshat and Kolla, Raja and Khatri, Chandra and Agarwal, Shubham},
  booktitle={CVPR 2025 Workshop Vision Language Models For All}
}

Contact

Contributions are welcome! If you have any improvements or suggestions, feel free to submit a pull request on GitHub.

Acknowledgement

Chitrakshara is built with reference to the code of the following projects: OBELICS. Thanks for their awesome work!