Datasets:

ananddey
/

asm-corpus

language: as
license: odc-by
task_categories:
  - text-generation
size_categories:
  - 1B-10B
tags:
  - assamese
  - pretraining
  - monolingual
  - asm_Beng
  - indic

AsmCorpus — Assamese Pretraining Dataset

The largest open monolingual Assamese corpus for LLM pretraining.

Documents: 2.37M
Characters: 11B
GPT-2 tokens: ~3.7B | Gemma 4 E2B tokens: ~5.8B
Format: Parquet (text column only)
License: ODC-By 1.0

Usage

from datasets import load_dataset

ds = load_dataset("ananddey/asm-corpus", split="train", streaming=True)
for doc in ds:
    print(doc["text"])

How It Was Built

All documents passed through language detection, quality filtering, PII redaction, and global deduplication before inclusion. See the source repository for the full pipeline code.

Quality Notes

~90% of volume is machine-translated encyclopedia content — more formal register than natural web text
No detectable Bengali contamination (GlotLID asm_Beng vs ben_Beng)
87% document retention through all quality filters

Citation

@misc{asm-corpus,
  author = {Anand Dey},
  title = {AsmCorpus: An Assamese Pretraining Dataset},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {https://huggingface.co/datasets/ananddey/asm-corpus},
}