Add dataset card and link to paper

README.md

@@ -0,0 +1,28 @@
+---
+task_categories:
+- text-generation
+---
+
+# Data Augmentations for Data-Constrained Language Model Pretraining
+
+This repository contains the dataset shards used in the paper "[Data Augmentations for Data-Constrained Language Model Pretraining](https://huggingface.co/papers/2606.16246)".
+
+The dataset consists of approximately 75 million tokens from **DCLM-RefinedWeb**, specifically curated to investigate data augmentation as a regularizer in data-constrained, multi-epoch pretraining regimes.
+
+## Links
+
+- **Paper:** [Data Augmentations for Data-Constrained Language Model Pretraining](https://huggingface.co/papers/2606.16246)
+- **GitHub Repository:** [michaelchen-lab/data-augmentations-for-pretraining](https://github.com/michaelchen-lab/data-augmentations-for-pretraining)
+
+## Dataset Summary
+
+The dataset was used to train a 150M-parameter Llama-based model for 100 epochs. It is organized into processed JSONL shards:
+- **Training data:** `shard_XXXXXXXX_processed.jsonl` files (approximately 75M tokens).
+- **Validation data:** `val_shard_00000000_processed.jsonl`.
+
+The researchers used this data to evaluate three categories of augmentations:
+1. **Token-level noise:** Masking or random token replacement.
+2. **Sequence permutations:** Right-to-left prediction and Fill-in-the-Middle (FIM).
+3. **Target offset prediction:** Predicting $x_{t+i}$ for $i > 1$.
+
+For detailed information on how to use these shards with the official pretraining scripts, please refer to the [GitHub repository](https://github.com/michaelchen-lab/data-augmentations-for-pretraining).