Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# WebOrganizer/Corpus-200B
|
| 2 |
+
[[Paper](ARXIV_TBD)] [[Website](https://weborganizer.allenai.org)] [[GitHub](https://github.com/CodeCreator/WebOrganizer)]
|
| 3 |
+
We pre-process the `1b-1x` pool from DataComps-LM using [RefinedWeb filters](https://github.com/mlfoundations/dclm/blob/main/baselines/baselines_configs/dclm_baseline_refinedweb.yaml) and [BFF deduplication](https://github.com/mlfoundations/dclm/tree/main/dedup/bff).
|
| 4 |
+
The resulting 200B token corpus is available here together with quality scores and domain annotations.
|
| 5 |
+
__Download the dataset by cloning the repository with Git LFS instead of HuggingFace's `load_dataset()`.__
|
| 6 |
+
The dataset has the following folder structure:
|
| 7 |
+
```bash
|
| 8 |
+
Corpus-200B/
|
| 9 |
+
documents/ # Pre-processed web documents
|
| 10 |
+
- CC_shard_00000000_processed.jsonl.zst
|
| 11 |
+
- CC_shard_00000001_processed.jsonl.zst
|
| 12 |
+
- ...
|
| 13 |
+
tokens/ # number of tokens per document
|
| 14 |
+
- CC_shard_00000000_processed.npy
|
| 15 |
+
- CC_shard_00000001_processed.npy
|
| 16 |
+
- ...
|
| 17 |
+
scores_dclm-fasttext/ # DCLM-fasttext score
|
| 18 |
+
- CC_shard_00000000_processed.npy
|
| 19 |
+
- ...
|
| 20 |
+
scores_fineweb-edu/ # FineWeb-Edu score
|
| 21 |
+
- CC_shard_00000000_processed.npy
|
| 22 |
+
- CC_shard_00000000_processed__rounded.npy # rounded FineWeb-Edu score
|
| 23 |
+
- ...
|
| 24 |
+
domains_topics/ # TopicClassifier annotations
|
| 25 |
+
- CC_shard_00000000_processed__logits.npy # logits for each topic
|
| 26 |
+
- CC_shard_00000000_processed__choice.npy # index of top choice
|
| 27 |
+
- ...
|
| 28 |
+
domains_formats/ # FormatClassifier annotations
|
| 29 |
+
- CC_shard_00000000_processed__logits.npy # logits for each format
|
| 30 |
+
- CC_shard_00000000_processed__choice.npy # index of top choice
|
| 31 |
+
- ...
|
| 32 |
+
domains_clusters-k24/ # K-means clusters
|
| 33 |
+
- CC_shard_00000000_processed.npy # cluster assignment for each document
|
| 34 |
+
- ...
|
| 35 |
+
```
|
| 36 |
+
We also include statistics about the presence and co-occurence of domains in the `domain_statistics/` folder, computed with the `domain_statistics.py` script.
|
| 37 |
+
## Citation
|
| 38 |
+
```bibtex
|
| 39 |
+
@article{wettig2025organize,
|
| 40 |
+
title={Organize the Web: Constructing Domains Enhances Pre-Training Data Curation},
|
| 41 |
+
author={Alexander Wettig and Kyle Lo and Sewon Min and Hannaneh Hajishirzi and Danqi Chen and Luca Soldaini},
|
| 42 |
+
year={2025}
|
| 43 |
+
}
|
| 44 |
+
```
|