AiventraLab/maithili-corpus

MaithiliBERT: First Pretrained Language Model for Maithili

MaithiliBERT is a RoBERTa-base model (125M parameters) pretrained on MaithiliCorpus — the largest open Maithili text corpus (39.8M words, LID-filtered and deduplicated).

This is the first pretrained language model specifically for Maithili, a low-resource Indic language spoken by ~34 million people.

Model Details

Property	Value
Architecture	RoBERTa-base
Parameters	124M
Training Data	MaithiliCorpus (gold_deduped, 39.8M words, 47.9K docs)
Vocabulary	BPE, 50K tokens
Context Length	512 tokens
Training	10 epochs, effective batch 32, FP16
License	CC-BY-4.0

Performance

Task	Metric	Score
MLM Perplexity	Hold-out set	TBD
Domain Classification	4-class accuracy	TBD

Usage

from transformers import pipeline, AutoModelForMaskedLM, AutoTokenizer

model = AutoModelForMaskedLM.from_pretrained("AiventraLab/maithili-roberta")
tokenizer = AutoTokenizer.from_pretrained("AiventraLab/maithili-roberta")

fill = pipeline("fill-mask", model=model, tokenizer=tokenizer)
results = fill("भगवान् <mask>क ध्यान")
for r in results:
    print(r["sequence"], r["score"])

Training Details

• Data: LID-filtered + deduplicated subset of MaithiliCorpus (47,882 documents)
• Tokenizer: ByteLevel BPE with 50K vocabulary trained on the corpus
• Optimizer: AdamW (lr=5e-4, β₁=0.9, β₂=0.98, ε=1e-6)
• Schedule: Linear warmup for 10% of steps, linear decay
• Batch: Effective batch size 32 (8 per GPU × 4 gradient accumulation steps)
• Mixed precision: FP16

Acknowledgments

The authors gratefully acknowledge AnK Komputing for providing the computational resources and infrastructure used to conduct the simulations/analyses presented in this work. We thank the team at AnK Komputing for their technical support and access to their high-performance computing services.

Citation

@article{maithili2026corpus,
  title={MaithiliCorpus: An Auditable 65.6M-Word Resource and a Validation Study of Raw-Text Morphological Rule Induction},
  author={[Authors]},
  journal={arXiv preprint},
  year={2026}
}

Limitations

• Pretrained on a single domain-imbalanced corpus (55.9% literary, 0.7% children's literature)
• May not generalize well to dialects or register variation
• BPE tokenizer may split rare Maithili words inefficiently

See Also

• MaithiliCorpus Dataset
• Corpus Documentation