---
task_categories:
- text-generation
---

# Data-Constrained Language Model Pretraining Dataset

This repository contains a dataset snapshot used in the paper [Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws](https://huggingface.co/papers/2606.06888).

The dataset consists of pre-tokenized `.pt` files containing packed GPT-2-tokenized sequences, derived from the DCLM (Data-Centric Language Modeling) corpus. These snapshots were prepared to study pretraining in data-constrained, compute-rich regimes.

- **Paper:** [Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws](https://huggingface.co/papers/2606.06888)
- **Code:** [https://github.com/yixinw-lab/dc_pretrain](https://github.com/yixinw-lab/dc_pretrain)

## Sample Usage

You can download this dataset snapshot using the `huggingface-cli`:

```bash
huggingface-cli download zhiwei555/dclm_data_200m \
  --repo-type dataset \
  --local-dir dclm_data_200m
```

*(Note: Replace `dclm_data_200m` with the specific dataset size or name if you are accessing a different variant.)*

## File Structure

The pre-tokenized files are typically organized as follows:
- `dclm_train.pt`: The training split.
- `dclm_val.pt`: The validation split.

These files are intended to be used with the training scripts provided in the official [GitHub repository](https://github.com/yixinw-lab/dc_pretrain).