agentlans/multilingual-text

Multilingual Text Dataset

This dataset contains a curated selection of rows from multiple input datasets, where each row includes a text chunk of approximately 2000 tokens (as measured by Llama 3.1 tokenizer) verified to be written in the correct language. Only rows with properly classified language chunks are retained, ensuring high-quality multilingual data for analysis or model training.

Preprocessing Steps

• Normalized whitespace, punctuation, Unicode characters, and bullet points
• Replaced sensitive information such as emails, phone numbers, and URLs with placeholders
• Language detection performed with the lid218e.bin FastText model
• Downloaded 20 000 chunks from each dataset
• Duplicate chunks were removed

Example Entry (mya_Mymr split)

{
  "uuid": "6d1baa56-f2d9-4122-ba9a-5ba518fbe199",
  "text": "်ပွားပြီး အိတ်ချ်အိုင်ဗွီရှိပြီး တီဘီကူးစက်ခံရသူ ၃၀% တွင် တီဘီရောဂါဖြစ်ပွားသည်။\nတီဘီရောဂါသည် လူဦးရေထူထပ်ခြင်းနှင့် အာဟာရချို့တဲ့ခြင်းများနှင့်လည်း ဆက်စပ်နေသည်။ ...",
  "language": "mya_Mymr",
  "dataset": "wikimedia/wikipedia",
  "config": "20231101.my"
}