Update README.md

README.md

@@ -1,3 +1,101 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+tags:
+- pretokenized
+- pretraining
+- llama3
+- text
+language:
+- en
+---
+
+### !! Note: this dataset is currently being uploaded and processed. The .bin files are intermediate files to allow shuffling. !!
+
+## DCLM-Baseline Pretokenized (LLaMA 3.1, 524288 context)
+
+This dataset is a pretokenized and globally shuffled version of DCLM-Baseline (mlfoundations/dclm-baseline-1.0), prepared for large-scale language model pretraining. It is intended to be used as a direct drop-in pretraining corpus for LLaMA 3.1 style training pipelines.
+
+The original DCLM-Baseline dataset is a 4T token / 3B document pretraining dataset that achieves strong performance on language model benchmarks. See the original dataset card for benchmark comparisons and further details about the source data.
+
+---
+
+### Tokenization
+
+Every document from DCLM-Baseline was tokenized using the LLaMA 3.1 tokenizer:
+
+- Tokenizer: meta-llama/Llama-3.1-8B
+- Library: HuggingFace transformers, AutoTokenizer
+- Each document is independently encoded via tokenizer.encode(document)
+- No padding, no truncation at the tokenization stage
+- Total documents: 2,949,254,346
+- Approximate total tokens: ~4T
+
+---
+
+### Global Shuffle
+
+A global document-level shuffle was applied across all 2,949,254,346 documents using a fixed random seed for full reproducibility:
+
+```python
+rng = np.random.default_rng(seed=42)
+permutation = np.arange(2949254346, dtype=np.int32)
+rng.shuffle(permutation)
+```
+
+This ensures that no locality from the original DCLM-Baseline shard ordering is retained in the final dataset.
+
+---
+
+### Sequence Packing
+
+Documents are packed into sequences of length 524,288 using a Best-Fit Cropping algorithm (as implemented in [karpathy/nanochat](https://github.com/karpathy/nanochat/blob/0aaca56805eb13f6e6e1fff789a08086902f12ab/nanochat/dataloader.py#L74-L161)). The key properties of this packing strategy are:
+
+- Every sequence starts with a BOS token
+- 100% token utilization, meaning no padding tokens are present
+- When no document fits in the remaining space, the shortest buffered document is cropped to fill exactly
+- Approximately 35% of tokens are discarded due to cropping at sequence boundaries
+
+---
+
+### Intended Use
+
+This dataset is intended purely for autoregressive language model pretraining. It is not suitable for fine-tuning, instruction tuning, or any supervised task. It inherits the research-only scope of the original DCLM-Baseline dataset.
+
+---
+
+### Out-of-Scope Use
+
+This dataset inherits all out-of-scope use statements from the original DCLM-Baseline dataset. It is not intended for production model training, and performance on domain-specific tasks such as code and math may be limited. It should be used for research purposes only.
+
+---
+
+### Source Data
+
+The underlying data is DCLM-Baseline, which is derived from Common Crawl web crawl data. Full details on curation, filtering, and deduplication are available in the original dataset card and the associated paper:
+
+- Dataset: https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0
+- Paper: https://arxiv.org/abs/2406.11794
+- Code: https://github.com/mlfoundations/dclm
+
+---
+
+### License
+
+CC-BY-4.0, inherited from the original DCLM-Baseline dataset.
+
+---
+
+### Citation
+
+If you use this dataset, please cite the original DCLM paper:
+
+```bibtex
+@misc{li2024datacomplm,
+      title={DataComp-LM: In search of the next generation of training sets for language models},
+      author={Jeffrey Li and Alex Fang and Georgios Smyrnis and Maor Ivgi and Matt Jordan and Samir Gadre and Hritik Bansal and Etash Guha and Sedrick Keh and Kushal Arora and Saurabh Garg and Rui Xin and Niklas Muennighoff and Reinhard Heckel and Jean Mercat and Mayee Chen and Suchin Gururangan and Mitchell Wortsman and Alon Albalak and Yonatan Bitton and Marianna Nezhurina and Amro Abbas and Cheng-Yu Hsieh and Dhruba Ghosh and Josh Gardner and Maciej Kilian and Hanlin Zhang and Rulin Shao and Sarah Pratt and Sunny Sanyal and Gabriel Ilharco and Giannis Daras and Kalyani Marathe and Aaron Gokaslan and Jieyu Zhang and Khyathi Chandu and Thao Nguyen and Igor Vasiljevic and Sham Kakade and Shuran Song and Sujay Sanghavi and Fartash Faghri and Sewoong Oh and Luke Zettlemoyer and Kyle Lo and Alaaeldin El-Nouby and Hadi Pouransari and Alexander Toshev and Stephanie Wang and Dirk Groeneveld and Luca Soldaini and Pang Wei Koh and Jenia Jitsev and Thomas Kollar and Alexandros G. Dimakis and Yair Carmon and Achal Dave and Ludwig Schmidt and Vaishaal Shankar},
+      year={2024},
+      eprint={2406.11794},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG}
+}
+```