Datasets:

HuggingFaceFW
/

fineweb

#53 opened 2 months ago by

#52 opened 3 months ago by

#50 opened 3 months ago by

#49 opened 3 months ago by

#48 opened 3 months ago by

#47 opened 3 months ago by

#46 opened 4 months ago by

#45 opened 4 months ago by

#44 opened 4 months ago by

#43 opened 4 months ago by

#42 opened 5 months ago by

#41 opened 5 months ago by

#40 opened 5 months ago by

#39 opened 5 months ago by

#38 opened 5 months ago by

#36 opened 5 months ago by

#35 opened 5 months ago by

#33 opened 5 months ago by

#32 opened 5 months ago by

#29 opened 5 months ago by

#28 opened 6 months ago by

#25 opened 6 months ago by

#19 opened 6 months ago by

#14 opened 7 months ago by

#12 opened 7 months ago by

#9 opened 7 months ago by

#8 opened 7 months ago by

#4 opened 7 months ago by

350BT sample is much smaller than advertized

Updates

Exact copy of this dataset on HuggingFace yields "This dataset has 218 files that have been marked as unsafe."

Simple exact deduplication removes 2/3 of data.

Individual files are too large to be used in some Parquet readers

Truly the finest of the web

requests.exceptions.HTTPError: 503 Server Error: Service Unavailable for url: https://huggingface.co/api/datasets/HuggingFaceFW/fineweb-edu/paths-info/5b89d1ea9319fe101b3cbdacd89a903aca1d6052

Dataset Viewer issue: JobManagerCrashedError

MMLU evaluation setting

Fineweb Download Size Discrepency

Is there an official test set for benchmarking objectively?

Any plans to release warc content after the language filtering steps?

Casting Issue?

Fineweb train configuration

Reproducibility of the work for other languages

why do you apply "All filters except the (very destructive) terminal_punct"

How to compute the aggerate score?

Language subset

Dedup

Regarding to the newly updated indexes(writen as deduplication issues)

"Reproducing GPT-2 (124M) in llm.c in 90 minutes for $20" using fineweb by Karpathy

Unsafe files

tiny-fineweb

Training configs for data ablation study

Reprocessing for a new language

Are copyrighted works included in this dataset?

Any plan to train models on larger subset of dataset?

Torrent?