MWirelabs/northeast-languages-test-set

Northeast Languages Test Set

A curated test set of 500 deduplicated sentences per language for evaluating language models on Northeast Indian languages.

Languages

This dataset contains test data for 9 Northeast Indian languages:

• Assamese (asm) - 500 sentences
• Garo (grt) - 500 sentences
• Khasi (kha) - 500 sentences
• Kokborok (trp) - 500 sentences
• Meitei (mni) - 500 sentences
• Mizo (lus) - 500 sentences
• Naga (nag) - 500 sentences
• Nyishi (njz) - 500 sentences
• Pnar (pbv) - 500 sentences

Total: 4,500 sentences

Dataset Construction

All test sentences were:

1. Extracted from newer corpora not present in NE-BERT training data
2. Deduplicated against the training corpus
3. Randomly sampled to ensure diversity

Usage

from datasets import load_dataset

# Load all languages
dataset = load_dataset("MWirelabs/northeast-languages-test-set")

# Load specific language
khasi_test = load_dataset("MWirelabs/northeast-languages-test-set", split="khasi")

Evaluation

This test set was used in the NE-BERT paper for perplexity evaluation.

Citation

@inproceedings{nebert2025,
  title={NE-BERT: A Multilingual Language Model for Nine Northeast Indian Languages},
  author={Anonymous},
  booktitle={Proceedings of LoResLM Workshop at EACL 2026},
  year={2026}
}

License

CC-BY-4.0