MahaSTS / README.md
l3cube-pune's picture
Update README.md
e0a6bba verified
metadata
language:
  - mr
license: cc-by-4.0
size_categories:
  - 10K<n<100K
task_categories:
  - sentence-similarity
  - text-retrieval
  - text-ranking
pretty_name: MahaSTS
tags:
  - Marathi NLP
  - Sentence Similarity
  - Marathi STS
  - low-resource

MahaSTS Dataset

The MahaSTS dataset is a human-annotated Sentence Textual Similarity (STS) dataset for Marathi, consisting of 16,860 sentence pairs labeled with continuous similarity scores in the range of 0-5. It is designed to enable effective training for sentence similarity tasks in Marathi, particularly in low-resource settings.

Paper: L3Cube-MahaSTS: A Marathi Sentence Similarity Dataset and Models
Code: https://github.com/l3cube-pune/MarathiNLP
Project page: https://github.com/l3cube-pune/MarathiNLP

Overview:

The MahaSTS Dataset is a human-annotated dataset for Sentence Textual Similarity (STS) in Marathi, designed to train and evaluate models on sentence similarity tasks. The dataset contains 16,860 Marathi sentence pairs, each labeled with a continuous similarity score in the range of 0–5. The dataset is split into training, validation, and test sets with a ratio of 85:10:5, ensuring balanced supervision.

Alongside the dataset, the MahaSBERT-STS-v2 model is fine-tuned for regression-based similarity scoring, providing a baseline for Marathi sentence similarity tasks.

Language:

  • Primary Language: Marathi (Low-resource Indic Language)

Dataset Size:

  • Total Sentence Pairs: 16,860
    • Train: 14,328 sentence pairs
    • Validation: 840 sentence pairs
    • Test: 1,692 sentence pairs
  • Bucket Distribution:
    • 6 similarity buckets (0-5)
    • 2,810 sentence pairs per bucket

Annotation:

Each sentence pair is labeled with a continuous similarity score in the range of 0 to 5. The labels represent the degree of similarity between the two sentences, with 0 indicating no similarity and 5 indicating high similarity.

Intended Use:

The dataset is intended for:

  • Sentence Similarity
  • Regression Tasks
  • Sentence Embeddings
  • Marathi Embedding Model Benchmarking

Model Benchmarks:

The MahaSBERT-STS-v2 model, fine-tuned on this dataset, provides a performance baseline. Other models like MahaBERT, MuRIL, IndicBERT, and IndicSBERT can be benchmarked for comparison.

Citation:

If you use this dataset, please cite the following paper:

@article{mirashi2025l3cube,
  title={L3Cube-MahaSTS: A Marathi Sentence Similarity Dataset and Models},
  author={Mirashi, Aishwarya and Joshi, Ananya and Joshi, Raviraj},
  journal={arXiv preprint arXiv:2508.21569},
  year={2025}
}

@article{joshi2022l3cube,
  title={L3cube-mahanlp: Marathi natural language processing datasets, models, and library},
  author={Joshi, Raviraj},
  journal={arXiv preprint arXiv:2205.14728},
  year={2022}
}

License

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).