Assamese Question Answering Dataset (AssameseQA)

Dataset Details

Dataset Description

The Assamese Question Answering Dataset (AssameseQA) is an extractive Question Answering (QA) dataset developed for training and evaluating multilingual transformer models on the Assamese language.

The dataset contains Assamese contexts, questions, and answers spanning multiple domains, including:

• Assamese history
• Geography
• Culture
• Education
• Science
• General knowledge
• Company policies
• Customer support
• Informational articles

The dataset was created to support research on low-resource language NLP and to improve Assamese Question Answering systems using multilingual transformer models such as mBERT, IndicBERT, MuRIL, and XLM-RoBERTa.

Approximately 50% of the dataset consists of manually created and verified QA pairs, while the remaining 50% is adapted from the Assamese split of AI4Bharat IndicQA after cleaning, normalization, and schema standardization.

Curated by: Sankhyahrick Swami

Language: Assamese (as)

Task: Extractive Question Answering

License: CC BY 4.0

---

Dataset Sources

Original Data

Approximately half of the dataset was manually created by the author.

The manually written portion includes original Assamese contexts and question-answer pairs covering:

• Customer support
• Company knowledge bases
• Assamese history
• Education
• Technology
• Science
• General knowledge
• Daily life

Every manually created QA pair was reviewed before inclusion.

---

External Data

Approximately half of the dataset is adapted from:

AI4Bharat IndicQA

https://huggingface.co/datasets/ai4bharat/IndicQA

Only the Assamese split was used.

The original dataset is licensed under CC BY 4.0.

The extracted samples were cleaned, normalized, and converted into a consistent CSV format.

---

Intended Uses

Direct Use

This dataset is suitable for:

• Fine-tuning mBERT
• Fine-tuning IndicBERT
• Fine-tuning MuRIL
• Fine-tuning XLM-RoBERTa
• Extractive Question Answering
• Retrieval-Augmented Generation (RAG)
• Assamese conversational AI
• Educational NLP research
• Low-resource language research

---

Out-of-Scope Use

This dataset is not intended for:

• Medical diagnosis
• Legal consultation
• Financial decision making
• Safety-critical applications

Models trained on this dataset should be evaluated before deployment in real-world applications.

---

Dataset Structure

The dataset is distributed in CSV format.

Repository files:

• train.csv
• test.csv
• combined_dataset.csv

Each row represents one Question Answering example.

Columns

Column	Description
Context	Assamese context or passage
Question	Assamese question
Answer	Correct answer
Starting Point	Character position where the answer begins inside the context
Verified	Indicates whether the sample has been manually verified

Example

Context	Question	Answer	Starting Point	Verified
অসমৰ ৰাজধানী দিছপুৰ।	অসমৰ ৰাজধানী কি?	দিছপুৰ	10	Yes

---

Dataset Creation

Motivation

Assamese remains a low-resource language for Question Answering.

This dataset was created to:

• Expand Assamese QA resources
• Support multilingual transformer research
• Improve Assamese NLP
• Facilitate mBERT fine-tuning
• Encourage open-source Assamese NLP research

---

Source Data

The dataset combines two different sources.

1. Manually Curated Data (~50%)

The manually created portion consists of original Assamese contexts and QA pairs written by the author.

Topics include:

• History
• Education
• Company documentation
• Customer support
• General knowledge
• Science
• Technology

Questions and answers were manually written to resemble natural Assamese usage.

---

2. AI4Bharat IndicQA (~50%)

The remaining portion comes from the Assamese split of AI4Bharat IndicQA.

The data was:

• cleaned
• normalized
• reformatted
• standardized
• merged with the manually created dataset

No machine translation was performed.

---

Data Processing

The manually created dataset underwent:

• Unicode normalization
• Duplicate removal
• Manual proofreading
• Assamese punctuation normalization
• Answer span verification
• CSV formatting

The AI4Bharat portion underwent:

• Schema conversion
• Unicode normalization
• Duplicate removal
• Formatting consistency checks

---

Source Data Producers

Original data:

• Sankhyahrick Swami

Derived data:

• AI4Bharat IndicQA contributors

---

Annotations

Annotation Process

For the manually created portion:

• Contexts were written manually.
• Questions were written manually.
• Answers were manually verified.
• Character-level answer positions were validated.

---

Annotators

Original annotations:

• Sankhyahrick Swami

Derived annotations:

• AI4Bharat IndicQA contributors

---

Personal and Sensitive Information

The dataset intentionally excludes:

• Personal information
• Contact details
• Financial records
• Medical records
• Private conversations

Most contexts consist of educational material, public information, or fictional company documentation.

---

Bias, Risks, and Limitations

The dataset may contain:

• Topic imbalance toward educational and customer-support domains
• Limited dialectal variation
• Limited conversational Assamese
• Fewer examples from highly technical domains

Since approximately half of the dataset is adapted from AI4Bharat IndicQA, users should also review the documentation and limitations of the original dataset.

---

Recommendations

This dataset is recommended for:

• Extractive QA
• Retrieval systems
• Multilingual NLP
• Assamese language modeling
• Academic research

For production systems, combining this dataset with additional Assamese corpora is recommended.

---

License

This dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Because part of the dataset is adapted from AI4Bharat IndicQA, users must comply with the attribution requirements of CC BY 4.0.

When using this dataset, please cite both:

• AssameseQA
• AI4Bharat IndicQA

---

Citation

If you use this dataset, please cite:

@dataset{AssameseQA2026,
  author = {Sankhyahrick Swami},
  title = {AssameseQA: An Assamese Question Answering Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/sankhyahrick/AssameseQA}
}

Please also cite the original AI4Bharat IndicQA dataset according to its official citation.

---

Acknowledgements

This dataset would not have been possible without the publicly available Assamese portion of the AI4Bharat IndicQA dataset. The manually created portion was developed as part of an internship project focused on multilingual Question Answering for Assamese.

---

Authors

Sankhyahrick Swami

---

Contact

For questions, suggestions, or bug reports, please open an issue on the Hugging Face repository:

https://huggingface.co/datasets/sankhyahrick/AssameseQA