siyah1/malayalam-dataset-llm

Malayalam Web Text (Cleaned)

A cleaned, deduplicated Malayalam text corpus built for training and fine-tuning Malayalam language models. Filtered from a large multilingual web-text collection down to high-quality Malayalam documents using script-ratio, length, and boilerplate-detection heuristics.

Dataset summary

• Language: Malayalam (ml)
• Rows kept: 2,631,750 (of 2,693,052 scanned — a 97.7% keep rate)
• Domain: General web text — news, blogs, informational sites, and other publicly crawled Malayalam-language pages
• Format: Parquet, single train split
• License: ODC-BY (Open Data Commons Attribution License)

Cleaning pipeline

Each document went through the following filters before being kept:

Step	Rule
Malayalam-script ratio	At least 60% of non-whitespace characters must be in the Malayalam Unicode block (U+0D00–U+0D7F)
Length bounds	Document text must be between 200 and 50,000 characters
Line-repetition ratio	At most 30% of lines in a document may be exact duplicates (filters nav menus, repeated category tags, cookie banners)
Deduplication	Exact-match dedup on whitespace-normalized, lowercased text
Text normalization	Unicode NFC normalization; collapsed excess blank lines and repeated spaces/tabs

This removes a lot of the common web-scrape noise — site navigation text, boilerplate footers, ad copy, and near-duplicate re-crawled pages — while keeping the genuine Malayalam prose content intact.

Schema

Field	Type	Description
text	string	Cleaned Malayalam document text
url	string	Source URL of the original page
timestamp	string	Crawl timestamp of the original page
source	string	Source crawl identifier (e.g. mC4, OSCAR)

How to load

Using 🤗 datasets

from datasets import load_dataset

ds = load_dataset("siyah1/malayalam-dataset-llm", split="train")
print(ds)
print(ds[0])

Streaming (recommended for large-scale use, avoids full download)

from datasets import load_dataset

ds = load_dataset("siyah1/malayalam-dataset-llm", split="train", streaming=True)

for example in ds.take(5):
    print(example["text"][:200])

Loading specific parquet shards directly

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files="https://huggingface.co/datasets/siyah1/malayalam-dataset-llm/resolve/main/data/train-*.parquet",
    split="train",
)

Using pandas (for a single shard / quick inspection)

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="siyah1/malayalam-dataset-llm",
    filename="data/train-00000-of-XXXXX.parquet",  # replace with actual shard filename
    repo_type="dataset",
)
df = pd.read_parquet(path)
df.head()

Example record

{
  'text': 'ഒരു Ketogenic ഡയറ്റ് എന്താണ്? ...',
  'url': 'https://ml.example.com/some-article/',
  'timestamp': '2021/04/20 13:11:47',
  'source': 'mC4',
}

Intended use

Suitable as text-generation / masked-language-modeling training data for Malayalam language models — e.g. building or adapting LLMs, tokenizer training, or general Malayalam NLP research. As with any web-derived corpus, downstream users should apply their own task-specific filtering or review before using it in production training runs.

Attribution

This is a derivative work distributed under ODC-BY, built from publicly available multilingual web-crawl data (mC4/OSCAR-derived). See the license field for terms.