Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws
Paper • 2606.06888 • Published
Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html>
<h"... is not valid JSON
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Data-Constrained Language Model Pretraining
This repository contains dataset snapshots used in the paper "Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws".
The dataset consists of pre-tokenized .pt files containing packed GPT-2-tokenized sequences derived from the DCLM corpus. These snapshots were curated to study language model pretraining in data-constrained, compute-rich regimes, specifically focusing on Masked-Input Regularization (MIR) and scaling laws.
Project Resources
- Paper: Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws
- GitHub Repository: https://github.com/yixinw-lab/dc_pretrain
Sample Usage
You can download the dataset snapshots using the huggingface-cli:
huggingface-cli download zhiwei555/dclm_data_200m \
--repo-type dataset \
--local-dir dclm_data_200m
The training scripts in the associated GitHub repository expect the following directory structure:
dclm_data_200m/dclm_train.ptdclm_data_200m/dclm_val.pt
Related Snapshots
The following snapshots were used for the experiments in the paper:
- 100M:
zhiwei555/dclm_data_100m - 200M:
zhiwei555/dclm_data_200m - 300M:
zhiwei555/dclm_data_300m - 400M:
zhiwei555/dclm_data_400m
- Downloads last month
- 65