README.md

metadata

annotations_creators:
  - manual
language_creators:
  - found
language:
  - en
  - te
  - kn
license: cc-by-4.0
multilinguality:
  - multilingual
pretty_name: Multilingual English-Telugu-Kannada Translation Dataset
size_categories:
  - 1M<n<10M
source_datasets:
  - ai4bharat/samanantar
task_categories:
  - translation
task_ids:
  - translation
dataset_info:
  features:
    - name: target_lang_code
      dtype: string
    - name: input_ids
      sequence: int32
    - name: attention_mask
      sequence: int8
    - name: labels
      sequence: int64
  splits:
    - name: train
      num_bytes: 2231633137
      num_examples: 7966485
  download_size: 700766001
  dataset_size: 2231633137
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

📚 Multilingual English-Telugu-Kannada Translation Dataset

This dataset is a curated and preprocessed subset of the AI4Bharat Samanantar dataset focused on multilingual translation tasks between English, Telugu (te_IN), and Kannada (kn_IN).

✨ Dataset Features

• Language pairs:
  • en ↔ te_IN
  • en ↔ kn_IN
• Preprocessed:
  • Filtered for sentence length (min=3, max=128 words)
  • Cleaned and normalized
• Tokenized using Hugging Face Transformers tokenizers:
  • M2M100Tokenizer (for en↔kn)
  • MBart50TokenizerFast (for en↔te)

📦 Dataset Structure

The dataset contains the following fields:

• src_texts: Source language sentence (English)
• tgt_texts: Target language sentence (Telugu or Kannada)
• labels: Tokenized target sequence for model training
• input_ids, attention_mask: Tokenized source sentence

The dataset is split into:

• train: Training samples
• validation: Small subset for evaluation

📊 Size

• ~7.9M total sentence pairs
• Supports batch training and multilingual fine-tuning

💡 Usage Example

from datasets import load_dataset

dataset = load_dataset("Koushim/en-te-kn-translation-dataset")

print(dataset["train"][0])

🧠 Intended Uses

• Train multilingual translation models (MBart, Marian, M2M100)
• Fine-tune LLMs on Indic translation
• Evaluate BLEU or other metrics for low-resource translation

📜 License

CC-BY-4.0

✍️ Author

Koushik Reddy GitHub | Hugging Face

🙏 Acknowledgements

Thanks to AI4Bharat for providing the Samanantar dataset which served as the base for this project.