Craw4LLM

This repo contains the code for the paper "Craw4LLM: Efficient Web Crawling for LLM Pretraining".

Prerequisite

1. Request the ClueWeb22 dataset.
2. Create a virtual environment with python >= 3.10 and install the following requirements:

numpy
tqdm
fasttext
pyyaml
wandb

3. Download the DCLM fastText classifier to fasttext_scorers/.

[!IMPORTANT] To run the crawler efficiently, the ClueWeb22 data should be placed on an SSD.

Run the Crawler

To run a (simulated) crawl, first create a yaml configuration file under configs/, and run the following command:

python crawl.py crawl --config 

Craw4LLM

Create a yaml file in configs/ with the following content:

cw22_root_path: 
seed_docs_file: seed.txt
output_dir: crawl_results/seed_10k_crawl_20m_dclm_fasttext
num_selected_docs_per_iter: 10000
num_workers: 16  # set to a number that fits your machine
save_state_every: -1  # set to a positive number to save the state (queue & visited set) of the crawler every certain steps
max_num_docs: 20000000
selection_method: dclm_fasttext_score
order: desc  # desc for descending, asc for ascending
wandb: true  # set to false to disable wandb logging
wandb_project: crawler
wandb_run_name: seed_10k_crawl_20m_dclm_fasttext
rating_methods:
    - 
        type: length
    - 
        type: fasttext_score
        rater_name: dclm_fasttext_score
        model_path: fasttext_scorers/openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train.bin

Documents are scored by all scorers in rating_methods. In the above configuration file, we set a length scorer, which scores a document by its length, and a fasttext_score scorer which uses the DCLM fastText model to score a document. The final ranking is determined by selection_method which is set to dclm_fasttext_score, the name of the fasttext_score scorer.

Baseline Crawlers

Random Crawler

cw22_root_path: 
seed_docs_file: seed.txt
output_dir: crawl_results/seed_10k_crawl_20m_random
num_selected_docs_per_iter: 10000
num_workers: 16
save_state_every: -1
max_num_docs: 20000000
selection_method: random_score
order: desc
wandb: true
wandb_project: crawler
wandb_run_name: seed_10k_crawl_20m_random
rating_methods:
    - 
        type: random_score

Indegree-based Crawler

cw22_root_path: 
seed_docs_file: seed.txt
output_dir: crawl_results/seed_10k_crawl_20m_indegree
num_selected_docs_per_iter: 10000
num_workers: 16
save_state_every: -1
max_num_docs: 20000000
selection_method: inlink_count
order: desc
wandb: true
wandb_project: crawler
wandb_run_name: seed_10k_crawl_20m_indegree
rating_methods:
    - 
        type: inlink_count

Pretraining and Evaluation

After running the crawler, the crawled document ids will be placed in output_dir in the configuration file. Run the following command to get the document texts:

python fetch_docs.py  --input_dir   --output_dir   --num_workers 

Then you can use the DCLM framework to run LLM pretraining and evaluation.

Miscellaneous

Browse the Data

Run the following command to print a document and its outlinks by its id:

python access_data.py