File size: 2,732 Bytes
8525949 420c3f5 5e90af9 7e6519b 420c3f5 7e6519b 420c3f5 7e6519b 420c3f5 669caf1 420c3f5 669caf1 420c3f5 669caf1 420c3f5 7e6519b 420c3f5 7e6519b 420c3f5 669caf1 420c3f5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 |
---
language:
- en
size_categories:
- 100B<n<1T
---
# WebOrganizer/Corpus-200B
[[Paper](https://arxiv.org/abs/2502.10341)] [[Website](https://weborganizer.allenai.org)] [[GitHub](https://github.com/CodeCreator/WebOrganizer)]
This dataset is a pre-processed version of the `1b-1x` CommonCrawl pool from DataComps-LM cleaned with
(1) [RefinedWeb filters](https://github.com/mlfoundations/dclm/blob/main/baselines/baselines_configs/dclm_baseline_refinedweb.yaml) and
(2) [BFF deduplication](https://github.com/mlfoundations/dclm/tree/main/dedup/bff).
We provide the resulting 200B token corpus annotated with two quality scores, WebOrganizer domains, and k-means scores.
__Download the dataset by cloning the repository with Git LFS instead of HuggingFace's `load_dataset()`.__
The dataset has the following folder structure:
```bash
Corpus-200B/
documents/ # Pre-processed web documents
- CC_shard_00000000_processed.jsonl.zst
- CC_shard_00000001_processed.jsonl.zst
- ...
tokens/ # number of tokens per document (GPT-NeoX tokenizer)
- CC_shard_00000000_processed.npy
- CC_shard_00000001_processed.npy
- ...
scores_dclm-fasttext/ # DCLM-fasttext score
- CC_shard_00000000_processed.npy
- ...
scores_fineweb-edu/ # FineWeb-Edu score
- CC_shard_00000000_processed.npy
- ...
scores_fineweb-edu__rounded/ # Rounded FineWeb-Edu score
- CC_shard_00000000_processed__rounded.npy
- ...
domains_topics/ # TopicClassifier annotations
- CC_shard_00000000_processed__choice.npy # index of top choice
- ...
domain_topics__logits/
- CC_shard_00000000_processed__logits.npy # logits for each topic
- ...
domains_formats/ # FormatClassifier annotations
- CC_shard_00000000_processed__choice.npy # index of top choice
- ...
domains_formats/ # FormatClassifier annotations
- CC_shard_00000000_processed__logits.npy # logits for each format
- ...
domains_clusters-k24/ # K-means clusters
- CC_shard_00000000_processed.npy # cluster assignment for each document
- ...
```
We also include statistics about the presence and co-occurence of domains in the `domain_statistics/` folder, computed with the `domain_statistics.py` script.
## Citation
If you make use of this pre-processed corpus in your work, please cite:
```bibtex
@article{wettig2025organize,
title={Organize the Web: Constructing Domains Enhances Pre-Training Data Curation},
author={Alexander Wettig and Kyle Lo and Sewon Min and Hannaneh Hajishirzi and Danqi Chen and Luca Soldaini},
journal={arXiv preprint arXiv:2502.10341},
year={2025}
}
``` |