| | ---
|
| | license: apache-2.0
|
| | task_categories:
|
| | - text-generation
|
| | language:
|
| | - en
|
| | tags:
|
| | - redpajama
|
| | - llm
|
| | - dataset-reproduction
|
| | - redpajama-1b
|
| | - redpajama-subset
|
| | - redpajama-weighted
|
| | - redpajama-sample
|
| | - natural-language-processing
|
| | size_categories:
|
| | - 100M<n<1B
|
| | pretty_name: RedPajama 1B Weighted Subset
|
| | ---
|
| |
|
| | # RedPajama-1B-Weighted
|
| |
|
| | A **canonical 1 Billion token weighted subset** of the [RedPajama-Data-1T](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T) dataset.
|
| |
|
| | ## Dataset Description
|
| |
|
| | This is a strict, downsampled version of the [RedPajama-10B-Weighted](https://huggingface.co/datasets/krisbailey/RedPajama-10B-Weighted) dataset. It maintains the **exact domain distributions** of the full 1T dataset, resized to a lightweight **1 Billion token** footprint.
|
| |
|
| | This dataset is ideal for:
|
| | - **Rapid Prototyping:** Train small models or debug pipelines in minutes rather than days.
|
| | - **Reference Baselines:** Use a standard, well-defined subset for comparative benchmarks.
|
| | - **Educational Use:** Explore the properties of large-scale pretraining data on consumer hardware.
|
| |
|
| | ## Dataset Details
|
| | - **Total Tokens:** ~1,000,000,000 (1 Billion)
|
| | - **Parent Source:** [krisbailey/RedPajama-10B-Weighted](https://huggingface.co/datasets/krisbailey/RedPajama-10B-Weighted)
|
| | - **Original Source:** [togethercomputer/RedPajama-Data-1T](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T)
|
| | - **Language:** English
|
| | - **Format:** Apache Parquet
|
| | - **Producer:** Kris Bailey
|
| |
|
| | ## Motivation
|
| | While the 10B subset is manageable, sometimes you need something even faster. A 1 Billion token dataset is the "Goldilocks" size for many initial experiments—large enough to train a statistically significant small language model (e.g., TinyLlama size) but small enough to download and process on a laptop.
|
| |
|
| | We created this by strictly downsampling the 10B dataset to ensure that the distribution remains consistent with the larger parent datasets.
|
| |
|
| | ## Dataset Creation Process
|
| |
|
| | ### 1. Source Selection
|
| | We utilized the `RedPajama-10B-Weighted` dataset as the source. This parent dataset was already constructed via weighted interleaving of the original RedPajama corpus.
|
| |
|
| | ### 2. Global Shuffling
|
| | The 10B dataset was globally shuffled (Seed: 43) to ensure that selecting the first $N$ tokens results in a random, representative sample, rather than a temporal slice.
|
| |
|
| | ### 3. Truncation
|
| | We selected the first **1 Billion tokens** from the shuffled stream.
|
| |
|
| | ### 4. Verification
|
| | We verified that the final subset retains the correct proportional mix of CommonCrawl, C4, GitHub, etc., matching the target distribution.
|
| |
|
| | ## Composition
|
| |
|
| | | Subset | Weight | Approx. Tokens |
|
| | | :--- | :--- | :--- |
|
| | | **CommonCrawl** | 74.16% | ~741.6 M |
|
| | | **C4** | 14.78% | ~147.8 M |
|
| | | **GitHub** | 4.98% | ~49.8 M |
|
| | | **ArXiv** | 2.36% | ~23.6 M |
|
| | | **Wikipedia** | 2.03% | ~20.3 M |
|
| | | **StackExchange** | 1.69% | ~16.9 M |
|
| |
|
| | ## Usage
|
| | ```python
|
| | from datasets import load_dataset
|
| |
|
| | # Load the 1B weighted subset
|
| | ds = load_dataset("krisbailey/RedPajama-1B-Weighted", split="train")
|
| |
|
| | print(ds)
|
| | ```
|
| |
|
| | ## Citation
|
| | If you use this dataset, please cite the original RedPajama work:
|
| | ```bibtex
|
| | @software{together2023redpajama,
|
| | author = {Together Computer},
|
| | title = {RedPajama: An Open Source Recipe to Reproduce LLaMA training dataset},
|
| | month = April,
|
| | year = 2023,
|
| | url = {https://github.com/togethercomputer/RedPajama-Data}
|
| | }
|
| | ```
|
| |
|
