Add dataset card, link to paper and code (#1)
Browse files- Add dataset card, link to paper and code (3a41bccac6f71609197631886ec85b1a5eacbdb7)
Co-authored-by: Niels Rogge <nielsr@users.noreply.huggingface.co>
README.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
task_categories:
|
| 3 |
+
- text-generation
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# Data-Constrained Language Model Pretraining
|
| 7 |
+
|
| 8 |
+
This repository contains pre-tokenized data used in the paper [Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws](https://huggingface.co/papers/2606.06888).
|
| 9 |
+
|
| 10 |
+
The dataset consists of pre-tokenized `.pt` files containing packed GPT-2-tokenized sequences. These snapshots were used to study the data-constrained, compute-rich regime of pretraining where models are trained for multiple epochs over a finite dataset.
|
| 11 |
+
|
| 12 |
+
- **Code:** [https://github.com/yixinw-lab/dc_pretrain](https://github.com/yixinw-lab/dc_pretrain)
|
| 13 |
+
- **Paper:** [https://huggingface.co/papers/2606.06888](https://huggingface.co/papers/2606.06888)
|
| 14 |
+
|
| 15 |
+
## Sample Usage
|
| 16 |
+
|
| 17 |
+
You can download the dataset snapshots used for these experiments using the `huggingface-cli`:
|
| 18 |
+
|
| 19 |
+
```bash
|
| 20 |
+
huggingface-cli download zhiwei555/dclm_data_200m \
|
| 21 |
+
--repo-type dataset \
|
| 22 |
+
--local-dir dclm_data_200m
|
| 23 |
+
```
|
| 24 |
+
|
| 25 |
+
The training scripts in the official repository expect these pre-tokenized `.pt` files as input.
|