| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - code |
| - python |
| - pretraining |
| - programming |
| - pip |
| - datasets |
| - gradio |
| - keras |
| - pytorch-lightning |
| - pytest |
| - transformers |
| - ternsorflow |
| - pytorch |
| - numpy |
| - pandas |
| - matpotlib |
| - django |
| size_categories: |
| - 100K<n<1M |
| pretty_name: Goodman |
| --- |
| |
| # Python and Docs Training Dataset |
|
|
| A massive, multi-format code and documentation corpus built from 70+ major open-source Python repositories (including the Pip, Keras, Gradio, Transformers, Datasets, and more). |
|
|
| ## Dataset Structure |
|
|
| The dataset is processed using exact BPE token chunking (**512-token chunks** with **zero overlap** via OpenAI's `cl100k_base` tokenizer) while preserving original Python indentation, newlines, and `<|file|>path` document boundaries—eliminating boundary duplication and allowing direct fixed-window pre-training without truncation or padding. |
|
|
| Each record in the JSONL and Parquet shards contains: |
| - **`id`**: Unique sequential identifier for the 512-token chunk (e.g., `combine_chunk_0`). |
| - **`source`**: Originating combined corpus file (`combine.txt`). |
| - **`chunk_index`**: Zero-based sequential index preserving repository-grouped ordering (`Docs -> Code`). |
| - **`token_count`**: Exact BPE token length (`512` tokens per row; `< 512` only on the final EOF tail row if retained). |
| - **`content`**: Raw, indentation-preserved block of Python code, type stubs, or documentation separated by `<|file|>path` headers. |
|
|
| *(An accompanying flat binary shard `Python_pretrain.bin` stores the same continuous token stream as headerless `uint32` token IDs for zero-copy `np.memmap` training.)* |
|
|
| ## Corpus Breakdown |
| - **Combined Repository Corpus (`combine.txt`)**: **22,551,967 lines** (~850.3 MB / ~220M+ `cl100k_base` tokens) curated across **74 permissive-license (MIT, Apache-2.0, BSD-3) repositories**, with line lengths capped at `999` characters and test/vendored directories excluded. |
| - **Documentation (`.md`, `.rst`)**: Repository guides, API references, and technical manuals grouped ahead of each library's source tree. |
| - **Python Source & Type Stubs (`.py`, `.pyi`)**: Production library implementations, core modules, and type annotations across foundational Python, data science, web, and ML frameworks (including CPython, PyTorch, TensorFlow, NumPy, SciPy, Pandas, Django, FastAPI/Starlette, and Transformers). |
|
|
| ## License |
| Distributed under the **Apache-2.0** open-source license. |
