Datasets:

Modalities:
Text
Formats:
parquet
Languages:
Turkish
Libraries:
Datasets
Dask
License:
cxrbon16's picture
Update README.md
8c9a640 verified
metadata
dataset_info:
  features:
    - name: url
      dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 56947576117
      num_examples: 9075453
  download_size: 22825493949
  dataset_size: 56947576117
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-4.0
language:
  - tr
pretty_name: c

This is the Turkish pretraining corpus of the Cosmos AI Research Group. It contains ~15B tokens and demonstrates competitive performance across various Turkish benchmarks when used in continual pretraining setups.

Cosmos-Turkish-Corpus is collected from a wide range of Turkish websites, including forums, news sources, Wikipedia, and more. URL-based deduplication has been applied; however, additional content-level deduplication and filtering may be required before use.