Tried and tested mixes for strong pretraining. Inspired by https://huggingface.co/blog/codelion/optimal-dataset-mixing
AI & ML interests
We release large pre-training datasets to accelerate open LLM development. Part of the Hugging Face Science team (hf.co/science)
Recent Activity
View all activity
Papers
FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language
The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
-
HuggingFaceFW/finepdfs
Viewer β’ Updated β’ 476M β’ 35.8k β’ 905 -
HuggingFaceFW/finepdfs-edu
Viewer β’ Updated β’ 49.5M β’ 4.5k β’ 95 -
HuggingFaceFW/ocr-annotations
Viewer β’ Updated β’ 1.62k β’ 62 β’ 19 -
HuggingFaceFW/finepdfs_lang_classification
Viewer β’ Updated β’ 3.08M β’ 14.9k β’ 4
-
FineWeb: decanting the web for the finest text data at scale
π·1.41kExplore and download the FineWeb webβscale text dataset
-
HuggingFaceFW/fineweb
Viewer β’ Updated β’ 52.5B β’ 374k β’ 3.13k -
HuggingFaceFW/fineweb-edu
Viewer β’ Updated β’ 3.5B β’ 375k β’ 1.25k -
HuggingFaceFW/fineweb-edu-score-2
Viewer β’ Updated β’ 13.9B β’ 17.8k β’ 89
1.8B models trained on 350BT to compare different pretraining datasets
-
HuggingFaceFW/ablation-model-fineweb-edu
Text Generation β’ 2B β’ Updated β’ 62 β’ 24 -
HuggingFaceFW/ablation-model-fineweb-v1
Text Generation β’ 2B β’ Updated β’ 29 β’ 14 -
HuggingFaceFW/ablation-model-refinedweb
Text Generation β’ 2B β’ Updated β’ 10 β’ 3 -
HuggingFaceFW/ablation-model-c4
Text Generation β’ 2B β’ Updated β’ 22 β’ 4
-
FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language
Paper β’ 2506.20920 β’ Published β’ 79 -
HuggingFaceFW/fineweb-2
Viewer β’ Updated β’ 4.48B β’ 81.2k β’ 854 -
Scaling FineWeb to 1000+ languages: Step 1: finding signal in 100s of evaluation tasks
π95Evaluate multilingual models using FineTasks
FineWeb-Edu datasets, classifier and ablation model
-
HuggingFaceFW/fineweb-edu
Viewer β’ Updated β’ 3.5B β’ 375k β’ 1.25k -
HuggingFaceFW/fineweb-edu-score-2
Viewer β’ Updated β’ 13.9B β’ 17.8k β’ 89 -
HuggingFaceFW/fineweb-edu-classifier
Text Classification β’ 0.1B β’ Updated β’ 12.9k β’ β’ 223 -
HuggingFaceFW/ablation-model-fineweb-edu
Text Generation β’ 2B β’ Updated β’ 62 β’ 24
Ablation models trained for our data experiments.
-
HuggingFaceFW/ablation-exp-textext-warc_trafilatura-28BT
Text Generation β’ 2B β’ Updated β’ 12 β’ 1 -
HuggingFaceFW/ablation-exp-textext-wet-28BT
Text Generation β’ 2B β’ Updated β’ 6 -
HuggingFaceFW/ablation-exp-fw-base_filtering-350BT
Text Generation β’ 2B β’ Updated β’ 9 -
HuggingFaceFW/ablation-exp-dedup-global_minhash-350BT
Text Generation β’ 2B β’ Updated β’ 10
Tried and tested mixes for strong pretraining. Inspired by https://huggingface.co/blog/codelion/optimal-dataset-mixing
-
HuggingFaceFW/finepdfs
Viewer β’ Updated β’ 476M β’ 35.8k β’ 905 -
HuggingFaceFW/finepdfs-edu
Viewer β’ Updated β’ 49.5M β’ 4.5k β’ 95 -
HuggingFaceFW/ocr-annotations
Viewer β’ Updated β’ 1.62k β’ 62 β’ 19 -
HuggingFaceFW/finepdfs_lang_classification
Viewer β’ Updated β’ 3.08M β’ 14.9k β’ 4
-
FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language
Paper β’ 2506.20920 β’ Published β’ 79 -
HuggingFaceFW/fineweb-2
Viewer β’ Updated β’ 4.48B β’ 81.2k β’ 854 -
Scaling FineWeb to 1000+ languages: Step 1: finding signal in 100s of evaluation tasks
π95Evaluate multilingual models using FineTasks
-
FineWeb: decanting the web for the finest text data at scale
π·1.41kExplore and download the FineWeb webβscale text dataset
-
HuggingFaceFW/fineweb
Viewer β’ Updated β’ 52.5B β’ 374k β’ 3.13k -
HuggingFaceFW/fineweb-edu
Viewer β’ Updated β’ 3.5B β’ 375k β’ 1.25k -
HuggingFaceFW/fineweb-edu-score-2
Viewer β’ Updated β’ 13.9B β’ 17.8k β’ 89
FineWeb-Edu datasets, classifier and ablation model
-
HuggingFaceFW/fineweb-edu
Viewer β’ Updated β’ 3.5B β’ 375k β’ 1.25k -
HuggingFaceFW/fineweb-edu-score-2
Viewer β’ Updated β’ 13.9B β’ 17.8k β’ 89 -
HuggingFaceFW/fineweb-edu-classifier
Text Classification β’ 0.1B β’ Updated β’ 12.9k β’ β’ 223 -
HuggingFaceFW/ablation-model-fineweb-edu
Text Generation β’ 2B β’ Updated β’ 62 β’ 24
1.8B models trained on 350BT to compare different pretraining datasets
-
HuggingFaceFW/ablation-model-fineweb-edu
Text Generation β’ 2B β’ Updated β’ 62 β’ 24 -
HuggingFaceFW/ablation-model-fineweb-v1
Text Generation β’ 2B β’ Updated β’ 29 β’ 14 -
HuggingFaceFW/ablation-model-refinedweb
Text Generation β’ 2B β’ Updated β’ 10 β’ 3 -
HuggingFaceFW/ablation-model-c4
Text Generation β’ 2B β’ Updated β’ 22 β’ 4
Ablation models trained for our data experiments.
-
HuggingFaceFW/ablation-exp-textext-warc_trafilatura-28BT
Text Generation β’ 2B β’ Updated β’ 12 β’ 1 -
HuggingFaceFW/ablation-exp-textext-wet-28BT
Text Generation β’ 2B β’ Updated β’ 6 -
HuggingFaceFW/ablation-exp-fw-base_filtering-350BT
Text Generation β’ 2B β’ Updated β’ 9 -
HuggingFaceFW/ablation-exp-dedup-global_minhash-350BT
Text Generation β’ 2B β’ Updated β’ 10