Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
License:
| license: other | |
| language: | |
| - en | |
| pretty_name: Pytorch Code 50K | |
| size_categories: | |
| - 10K<n<100K | |
| tags: | |
| - code | |
| - python | |
| - pytorch | |
| - deep-learning | |
| - beta | |
| task_categories: | |
| - text-generation | |
| # ⚠️ Pytorch-Code-50K — BETA | |
| > **Status: BETA.** This is an early, unpolished release. The data is real but | |
| > **unaudited**: it has not been manually reviewed, re-verified against upstream, | |
| > or checked for toxic/low-quality content beyond the filters listed below. The | |
| > sample may change between releases. **Use at your own risk and do not treat it | |
| > as a benchmark.** | |
| ## Overview | |
| **Pytorch-Code-50K** is a collection of **~50,000 Python source files that use the | |
| PyTorch ecosystem**, extracted from BigCode's | |
| [**The Stack**](https://huggingface.co/datasets/bigcode/the-stack) | |
| (Python subset). | |
| It is intended as a lightweight, easy-to-consume corpus for: | |
| - fine-tuning and pre-training small code models on PyTorch idioms, | |
| - studying how PyTorch APIs are actually used in the wild, | |
| - retrieval / RAG over real PyTorch examples, | |
| - quick prototyping where a full 3 TB dataset is overkill. | |
| ## Source | |
| | | | | |
| |---|---| | |
| | **Upstream** | [bigcode/the-stack](https://huggingface.co/datasets/bigcode/the-stack) (`data/python`) | | |
| | **Curated by** | Monster-Code | | |
| | **Language** | Python | | |
| | **Domain** | PyTorch / deep learning code | | |
| | **Size** | ~50,000 files | | |
| | **Format** | Parquet (one row per file) | | |
| | **Release** | BETA v0.1 | | |
| > The Stack is a collection of source code from repositories with various | |
| > licenses. Any use of all or part of the code in this dataset must abide by the | |
| > terms of the original licenses, including attribution clauses when relevant. | |
| > Provenance (`repo`, `path`, `license`, `sha`) is preserved so you can | |
| > attribute and filter correctly. | |
| ## Selection criteria | |
| A file is included when **all** of the following hold: | |
| 1. It comes from the Python subset of The Stack. | |
| 2. Its source contains a top-level import from the PyTorch ecosystem, matched | |
| with this regex: | |
| ```python | |
| ^[ \t]*(?:import|from)[ \t]+( | |
| torch(vision|audio|text|arrow|metrics|_geometric|_scatter|_sparse| | |
| _cluster|_spline_conv|_xla|_cpp|_jit|_onnx|_fx)? | |
| | pytorch_lightning | |
| | lightning(\.pytorch)? | |
| | skorch | |
| )\b | |
| ``` | |
| 3. Its length is between **200 and 200,000 characters**. | |
| 4. Its content hash is unique within the sample (exact duplicates removed). | |
| ### Sampling | |
| The Stack stores files grouped by repository, so a naive "first 50K" pass would | |
| be dominated by whichever repositories sort first. Instead, `update.py` | |
| **reservoir-samples** over a randomly shuffled subset of the source shards, so | |
| the sample is approximately uniform over the scanned shards rather than biased | |
| toward any one repository. Pass `--seed` to reproduce or vary the sample. | |
| ## Dataset schema | |
| | Column | Type | Description | | |
| |---|---|---| | |
| | `content` | `string` | The full source code of the file. | | |
| | `repo` | `string` | Repository identifier (`owner/name`) when available. | | |
| | `path` | `string` | File path within the repository, when available. | | |
| | `license` | `string` | Detected license(s), comma-separated if multiple. | | |
| | `size` | `int64` | File size in bytes. | | |
| | `sha` | `string` | Content hash from the source dataset. | | |
| | `lang` | `string` | Language tag (always `python` here). | | |
| | `ext` | `string` | File extension (`.py`). | | |
| > Source column names vary between The Stack releases; `update.py` maps them | |
| > defensively. Columns that are missing upstream are emitted as empty strings. | |
| ## Usage | |
| ```python | |
| from datasets import load_dataset | |
| ds = load_dataset("Monster-Code/Pytorch-Code-50K", split="train") | |
| print(ds) | |
| # Dataset({ | |
| # features: ['content', 'repo', 'path', 'license', 'size', 'sha', 'lang', 'ext'], | |
| # num_rows: 50000 | |
| # }) | |
| # Only permissively licensed files | |
| ds = ds.filter(lambda x: "mit" in x["license"].lower()) | |
| # A single training example | |
| print(ds[0]["content"][:500]) | |
| ``` | |
| Stream it instead of downloading everything: | |
| ```python | |
| ds = load_dataset("Monster-Code/Pytorch-Code-50K", split="train", streaming=True) | |
| for row in ds: | |
| ... | |
| ``` | |
| ## How it was built | |
| ```bash | |
| pip install huggingface_hub pyarrow | |
| # accept the terms of the source dataset once: | |
| # https://huggingface.co/datasets/bigcode/the-stack | |
| python update.py # full 50K build | |
| python update.py --dry-run # scan and report, push nothing | |
| ``` | |
| `update.py`: | |
| 1. lists the Python parquet shards of `bigcode/the-stack`, | |
| 2. shuffles them with a seed and downloads them through the HF cache | |
| (re-runs are incremental), | |
| 3. filters by the PyTorch import regex and length bounds, | |
| 4. reservoir-samples 50,000 unique files, | |
| 5. writes local parquet shards and pushes them plus this card to the Hub. | |
| The script is idempotent and resumable: source shards are cached under `cache/` | |
| and already-downloaded shards are skipped. | |
| ## Known limitations (BETA) | |
| - **Not exhaustively filtered.** Files may contain security-sensitive code, | |
| secrets committed to version control, or license-incompatible snippets. No | |
| manual review has been performed. | |
| - **Import-based selection.** Files that use PyTorch only through a wrapper | |
| (e.g. `transformers` without importing `torch`) are **not** included. Files | |
| that mention `import torch` in a docstring/comment may be included. | |
| - **Sampling bias.** The sample is uniform over a random subset of shards, not | |
| over all PyTorch files on GitHub. Repositories whose code is heavily | |
| duplicated may still be over-represented. | |
| - **Duplicates.** The Stack is not deduplicated, so the same file can appear | |
| more than once (across repositories) and very similar files are common. | |
| - **License metadata is noisy.** `license` comes from upstream detection and | |
| may be missing, wrong, or list several licenses. | |
| - **Beta may change.** Rows and schema may change in future releases; pin a | |
| revision if you need reproducibility. | |
| ## Terms of use | |
| This dataset is distributed for research purposes. You are responsible for | |
| complying with the licenses of the original repositories. By using this dataset | |
| you agree to the [Terms of Use of The Stack](https://huggingface.co/datasets/bigcode/the-stack#terms-of-use-for-the-stack). | |
| ## Citation | |
| If you use this dataset, please cite The Stack and this release: | |
| ```bibtex | |
| @misc{pytorch-code-50k, | |
| title = {Pytorch-Code-50K (BETA)}, | |
| author = {Monster-Code}, | |
| year = {2026}, | |
| howpublished = {\url{https://huggingface.co/datasets/Monster-Code/Pytorch-Code-50K}} | |
| } | |
| @article{BigCodeTheStack, | |
| title = {The Stack: 3 TB of permissively licensed source code}, | |
| author = {BigCode Project}, | |
| journal = {arXiv preprint arXiv:2211.15533}, | |
| year = {2022} | |
| } | |
| ``` | |
| ## Contact | |
| Questions, takedown requests, or complaints about specific files: | |
| open a discussion on the dataset repo or contact the Monster-Code maintainers. | |
| --- | |
| *This is a **BETA** release. Expect rough edges. Please report issues on the | |
| dataset's Hugging Face page.* | |