BHAID: Bharatiya AI-generated Text Detection

BHAID is a multilingual benchmark dataset for detecting human-written and LLM-generated text in Indian languages. The dataset covers nine Indian languages: Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Odia, and Assamese.

BHAID is introduced in the paper:

BHAID: A Dataset of LLM-Generated Text in Indian Languages with a Dual Contrastive Learning Framework for Detection

The accompanying work proposes DualCLF, a Dual Hierarchical Contrastive Learning framework that combines self-supervised and supervised contrastive learning objectives with focal loss to improve the robustness of AI-generated text detection across Indian languages.

Dataset Overview

Property Details

---

Dataset BHAID Full name Bharatiya AI-generated Text Detection Total samples 27,476 Languages 9 Task Human vs. AI-generated text detection Data format Parquet Fields Text, Label, Lang

Languages

The dataset contains text from the following nine Indian languages:

Language Code

---

Hindi Hi Bengali Bn Tamil Ta Telugu Te Kannada Kn Malayalam Ml Gujarati Gu Odia Or Assamese As

Dataset Format

The dataset is provided in final_lang.parquet.

Each example contains three fields:

Field Description

---

Text The text sample Label Class label: human or ai Lang Language code

Example:

Text:   "भारत के सामुद्रिक रास्तों की खोज..."
Label:  human
Lang:   Hi

Labels

The Label field contains two classes:

• human --- human-written text
• ai --- LLM-generated text

Dataset Construction

BHAID was developed to address the limited availability of multilingual resources for AI-generated text detection in Indian languages. The dataset includes human-written and LLM-generated text spanning nine languages and is intended to support evaluation of multilingual AI-text detection systems, particularly in low-resource and morphologically rich language settings.

The dataset accompanies research investigating supervised, contrastive, and multilingual approaches to AI-generated text detection.

Intended Use

BHAID can be used for:

• Human vs. LLM-generated text classification
• Multilingual AI-generated text detection
• Evaluation of cross-lingual detection models
• Contrastive-learning research
• Robustness studies for LLM-generated text detection
• Research on AI-text detection in Indian languages
• Benchmarking multilingual NLP models

Citation

If you use BHAID in your research, please cite the associated paper:

@inproceedings{bhaID,
  title     = {BHAID: A Dataset of LLM-Generated Text in Indian Languages with a Dual Contrastive Learning Framework for Detection},
  author    = {Anonymous},
  booktitle = {Proceedings of the Annual Meeting of the Association for Computational Linguistics},
  year      = {2026}
}

Note: The author information above is intentionally anonymous while the associated paper is under anonymous review. Please replace the citation with the final published citation after acceptance.

Data Statistics

The current release contains 27,476 text samples across nine Indian languages.

Detailed language-wise and label-wise statistics can be added here:

Language Human AI Total

---

Hindi --- --- --- Bengali --- --- --- Tamil --- --- --- Telugu --- --- --- Kannada --- --- --- Malayalam --- --- --- Gujarati --- --- --- Odia --- --- --- Assamese --- --- --- Total --- --- 27,476

Loading the Dataset

Using 🤗 Datasets:

from datasets import load_dataset

dataset = load_dataset(
    "Sifi-world/BHAID",
    data_files="final_lang.parquet"
)

print(dataset)

Access the training split:

data = dataset["train"]

print(data[0])

Convert to pandas:

df = dataset["train"].to_pandas()

print(df.head())
print(df["Label"].value_counts())
print(df["Lang"].value_counts())

Data License

Please check the source and redistribution conditions of the underlying text before using the dataset for commercial or redistribution purposes.

The repository license should be updated to the exact license under which BHAID is released.

Limitations

BHAID is designed primarily as a benchmark for AI-generated text detection in Indian languages. Performance on this dataset should not be interpreted as evidence that a detector can reliably identify AI-generated text in all domains, languages, models, or real-world settings.

AI-generated text detection is sensitive to the generation model, prompting strategy, domain, language, text length, and post-processing. Users should therefore evaluate models under settings appropriate to their intended application.

Acknowledgements

This dataset is part of research on multilingual NLP and AI-generated text detection, with a focus on Indian languages and low-resource language settings.

Contact

For questions regarding the dataset, please use the Community tab of this Hugging Face repository.