README.md

metadata

license: apache-2.0
language:
  - en
task_categories:
  - text-generation
tags:
  - pretraining-data
  - data-curation
  - web-scraping
size_categories:
  - 10M<n<100M

ReScraper-Data

Data released with ReScraper, a 0.6B refiner that replaces the heuristic HTML-to-text stack of a pretraining data pipeline with a single small language model.

• Model: cx-cmu/ReScraper
• Code: cxcscmu/ReScraper

All four subsets come from the same source pool: the 10% raw-HTML sample of the DCLM pool (dclm-pool-400m-1x-html-jsonl-step3a-10pct), 18,025,558 pages over 10,319 shards.

Contents

• Subsets
• Schema
• Loading
• License

Subsets

Path	Rows	Size	What it is
data/ReScraper-Corpus/	10,056,105	12.4 GB	The curated pretraining corpus: model output after the post-filter and Bloom-filter deduplication. 7.44B GPT-NeoX-20B tokens.
data/ReScraper-Pool-Output/	10,646,236	13.3 GB	The refiner's raw output over the whole pool, before deduplication. Use this to study the model's behaviour page by page.
data/ReScraper-SFT-Stage1/	1,383,115	3.9 GB	Stage 1 training set: extraction plus the operation decision.
data/ReScraper-SFT-Stage2/	131,484	0.4 GB	Stage 2 training set, which upweights <rewrite> so the model learns that operation.

Files are Parquet, split into roughly 1 GB parts named <subset>-en-part-XXXX-of-YYYY.parquet.

Schema

Corpus and pool output

Field	Type	Meaning
text	string	the cleaned page
operation	string	the operation the model chose: <keep>, <edit> or <rewrite> (deleted pages are not materialized)
bff_contained_ngram_count_before_dedupe	int	deduplication bookkeeping (corpus only)

SFT sets

Field	Type	Meaning
input	string	the rendered page, one block per line with <lid:n> line ids
output	string	the target program: an <extract> block of line removals followed by the operation tag

Loading

from datasets import load_dataset

corpus = load_dataset("cx-cmu/ReScraper-Data", data_dir="data/ReScraper-Corpus", split="train")
sft    = load_dataset("cx-cmu/ReScraper-Data", data_dir="data/ReScraper-SFT-Stage1", split="train")

To reproduce pretraining, tokenize the corpus with the GPT-NeoX-20B tokenizer; the code repository has the tokenization and pretraining scripts.

License

Apache 2.0. The underlying pages come from Common Crawl via the DCLM pool and keep their original terms.