license: apache-2.0
English CC-News 2016-2021 cleaned, deduplicated, and decontaminated.
Cleaning
- Documents with non-English content are removed
- Other low-quality content such as advertisements are heuristically removed
Deduplication
- global deduplication within the dataset
- cross-deduplication with OpenWebText2, removing about 0.3M documents.
Decontamination
This dataset has been decontaminated with respect to the following benchmarks based on n-gram overlap:
- GLUE (dev set of SST-2, CoLA, QQP, WNLI, RTE, QNLI, MNLI; test set of MPRC)
- SIQA, PIQA, QASC, CSQA, HellaSWAG (all dev set)
- CONLL 2003
- BLIMP
- MAIN
- BoolQ (dev set)
- WinoGrande (dev set)
- ANLI (test set)
- ARC easy and challenge (test set)
- RACE middle and high (test set)
- MMLU (dev, val, and test sets)
- MATH, GSM8K (test set)
- HumanEval (test set)
- GPQA (diamond)
About 16K documents are removed in this step.
Dataset Statistics
Number of samples: 133,969,328.
Size of downloaded parquet files: 221G.
Filtered Version
There is a model-filtered version in the filtered branch, including 132,214,420 samples (219G parquet files).
Qwen2.5-32B-Instruct is used to generate language quality annotation (on a scale of 1-5) for 250K C4 samples. A RoBERT-large classifier is trained with regression on these annotations. Any document receiving a score of 1 or 2 from the classifier is removed. The remaining documents are also accompanied by their scores.
You can download this version by specifying the --revision argument:
huggingface-cli download --repo-type dataset Geralt-Targaryen/CC-News --revision filtered --local-dir .