| --- |
| language: |
| - en |
| pretty_name: NeoCorpus |
| configs: |
| - config_name: foundation |
| data_dir: foundation |
| - config_name: specialized |
| data_dir: specialized |
| task_categories: |
| - text-generation |
| tags: |
| - pretraining |
| - language-modeling |
| - english |
| license: cc-by-nc-4.0 |
| --- |
| # NeoCorpus |
|
|
| **NeoCorpus** is a multi-stage pretraining corpus for the NeoLM |
| family of decoder-only language models trained from scratch. |
|
|
| ## Available Configurations |
|
|
| | Configuration | Description | Status | |
| |---|---|---| |
| | `foundation` | General-purpose foundation pretraining | Available | |
| | `specialized` | Specialized continued pretraining | Planned | |
| | `decay` | Final high-quality pretraining | Planned | |
|
|
| ## Foundation Dataset |
|
|
| The `foundation` configuration contains an English-language subset |
| derived from Andrej Karpathy's globally shuffled NVIDIA Nemotron-ClimbMix. |
|
|
| ### Dataset Statistics |
|
|
| | Split | Documents | |
| |---|---:| |
| | Train | 1,889,952 | |
| | Validation | 5,000 | |
| | Test | 20,000 | |
| | **Total** | **1,914,952** | |
|
|
| ### Token Statistics |
|
|
| Estimated using the previous NeoLM ByteLevel BPE tokenizer |
| (vocabulary size 32,000), including one planned EOS per document. |
|
|
| | Split | Estimated Tokens | |
| |---|---:| |
| | Train | 1,199,827,981 | |
| | Validation | 3,223,381 | |
| | Test | 12,650,844 | |
|
|
| Final token counts will be recalculated using the tokenizer |
| trained for the next NeoLM model. |
|
|
| ### Source and Provenance |
|
|
| - Original corpus: [NVIDIA Nemotron-ClimbMix](https://huggingface.co/datasets/nvidia/Nemotron-ClimbMix) |
| - Shuffled source: [karpathy/climbmix-400b-shuffle](https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle) |
| - Source revision: `915333b4f8b8684f39aeaafea600fea6f43fb703` |
| - Collection seed: `1337` |
| - Collection method: sequential traversal of globally shuffled source shards |
|
|
| NeoCorpus is an independent derivative dataset, not affiliated |
| with NVIDIA or Andrej Karpathy. |
|
|
| ### Data Preparation |
|
|
| 1. Stream documents from the pinned source revision. |
| 2. Strip edge whitespace while retaining internal formatting. |
| 3. Filter empty documents and lengths outside 100–100,000 characters. |
| 4. Exclude reserved tokenizer markers. |
| 5. Reserve 5,000 validation and 20,000 test documents. |
| 6. Collect approximately 1.2 billion training tokens. |
| 7. Remove exact-text duplicates from the training set, |
| prioritizing validation and test holdouts. |
| 8. Preserve provenance and estimated token counts. |
|
|
| 483 training documents were removed by exact deduplication. |
| Near-duplicate deduplication was not performed. |
|
|
| ### Features |
|
|
| | Field | Description | |
| |---|---| |
| | `doc_id` | Stable document identifier | |
| | `text` | Document content | |
| | `source` | Original dataset identifier | |
| | `source_shard` | Original source shard | |
| | `token_count` | Estimated tokens including planned EOS | |
|
|
| ## Usage |
|
|
| from datasets import load_dataset |
| |
| ds = load_dataset( |
| "MaagDeveloper/NeoCorpus", |
| "foundation", |
| split="train", |
| ) |
| |
| The next NeoLM tokenizer will be trained on the training split only. |
|
|
| ## Limitations |
|
|
| The dataset may contain factual errors, biases, unsafe text, |
| personally identifiable information, and near-duplicates. |
| It has not undergone exhaustive manual quality review. |
|
|
| Source licensing and redistribution obligations require review |
| before public release. |
|
|
| ## Licensing and Redistribution |
|
|
| NeoCorpus is an independently curated derivative of [NVIDIA Nemotron-ClimbMix](https://huggingface.co/datasets/nvidia/Nemotron-ClimbMix), obtained through [Andrej Karpathy's globally shuffled redistribution](https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle). |
|
|
| The original NVIDIA dataset is published under the **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)** license. NeoCorpus is redistributed subject to the applicable attribution and non-commercial conditions of that license. |
|
|
| The following modifications were made: |
| - Selection of a subset from globally shuffled source shards. |
| - Document length filtering and whitespace normalization. |
| - Removal of exact duplicate training documents. |
| - Creation of separate training, validation, and test splits. |
| - Addition of document identifiers, provenance metadata, and estimated token counts. |
|
|
| The original NVIDIA dataset and Karpathy's redistribution are credited in the Attribution section below. |
|
|
| NeoCorpus is an independent, non-commercial research project and is not officially affiliated with NVIDIA or Andrej Karpathy. |
|
|
| This licensing notice does not grant additional rights over any third-party material contained in the source corpus. Users are responsible for complying with applicable licenses and rights. |
|
|
| ## Attribution |
|
|
| NeoCorpus builds upon NVIDIA Nemotron-ClimbMix and |
| Andrej Karpathy's shuffled ClimbMix dataset used for nanochat. |
|
|
| - [NVIDIA Nemotron-ClimbMix](https://huggingface.co/datasets/nvidia/Nemotron-ClimbMix) |
| - [Karpathy ClimbMix Shuffle](https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle) |
| - [nanochat](https://github.com/karpathy/nanochat) |
|
|
| This project is independent and is not officially affiliated |
| with NVIDIA or Andrej Karpathy. |
|
|
| ## Temporary Specialized Configuration |
|
|
| **WARNING: TEST DATA — NOT FOR MODEL TRAINING.** |
|
|
| The `specialized` configuration currently contains only 10 |
| synthetic, original short stories per split. |
|
|
| These 30 documents are temporary placeholders used to verify |
| multi-configuration behavior in the Hugging Face Dataset Viewer |
| and the `datasets` Python library. |
|
|
| They will be replaced with the real specialized corpus. |
| The `token_count` field is null for these placeholder rows. |
|
|
