README.md

---
license: mit
language:
- en
- es
task_categories:
- text-generation
- fill-mask
size_categories:
- 10K<n<100K
pretty_name: PyMini
tags:
- DS Mini
- Pixel Datasets
- python
- code
- education
- small-models
- fine-tuning
- synthetic
---

# PyMini – A Compact Python Instruction Dataset for Small Language Models

**PyMini** is a synthetic instruction dataset built by [Inserloft](https://inserloft.com) to teach Python programming to small language models (<2B parameters). It covers fundamental Python concepts through a diverse set of tasks — code prediction, bug fixing, function completion, explanation, and conceptual Q&A. Available in **English, Spanish, or bilingual** versions.

## Dataset Description

PyMini provides a **high-quality, compact, and fully synthetic** dataset for fine-tuning models that need a solid grasp of Python without unnecessary complexity. Every example is generated programmatically, avoiding license issues and giving exact control over difficulty and coverage.

- **Languages:** English (`en`), Spanish (`es`), or mixed (`both`)
- **Number of examples:** Configurable; default release contains **~49,200 examples** (balanced across task types)
- **Format:** Parquet (native), auto-converted from the original generation output
- **Generated with:** Python script (no web scraping, no copyrighted code)

## Dataset Structure

Each example is a JSON object with three fields:

```json
{
  "instruction": "string (the task description)",
  "input": "string (code snippet or empty)",
  "output": "string (the expected answer/code/explanation)"
}
```

Example (English):

```json
{
  "instruction": "What is the output of the following code?",
  "input": "```python\nprint(3 + 4 * 5)\n```",
  "output": "23"
}
```

The `input` field may be empty for tasks like "Write a function...".

## Task Types and Distribution

| Task Type | Description | Approx. Weight |
|---|---|---|
| Predict Output | Given a code snippet, predict what it prints or evaluates to. | 25% |
| Fix Bug | Correct syntax or logical errors in provided code. | 15% |
| Fill Missing Line | Complete a partially written function body to fulfill its purpose. | 15% |
| Write Function | Write a whole function from a natural language specification. | 20% |
| Explain Code | Explain in plain language what a piece of code does. | 15% |
| Concept Question | Answer a theoretical question about Python (e.g., "What is a list?"). | 10% |
| Convert Code | Transform a loop into a comprehension or vice versa. | 5% |

All code examples are self-contained and focus on core Python: variables, types, conditionals, loops, functions, lists, dictionaries, sets, file handling, exceptions, basic OOP, and comprehensions.

## Usage

Load the dataset directly with the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("inserloft/PyMini", split="train")
print(dataset[0])
```

The dataset is stored natively in **Parquet** format, so it loads quickly and works out of the box with the `datasets` library, Dataset Viewer, and any Parquet-based tooling (pandas, DuckDB, Polars, etc.):

```python
import pandas as pd

df = pd.read_parquet("hf://datasets/inserloft/PyMini/train.parquet")
print(df.head())
```

## Data Fields

- **instruction** (str): Task description in the chosen language.
- **input** (str): Optional code block (formatted with ` ```python ` fences) or empty string.
- **output** (str): Expected response (may contain code blocks, explanations, or short answers).

## Data Splits

The default release includes a single training split. For evaluation, we recommend either:

- Holding out a random 5–10% of the data, or
- Using a separate, handcrafted test set of real-world Python problems.

## Generation Process

PyMini is entirely synthetic, generated by a Python script that randomly combines templates and safe code evaluation to produce diverse examples. Key aspects of the generation:

- **Deduplication:** A hash-based deduplication step prevents exact duplicate examples.
- **Reproducibility:** The script uses a fixed random seed (42) for reproducible generation.
- **Safety:** All code snippets are evaluated in a restricted environment (no file system access, limited built-ins).
- **Customization:** Adjust the number of examples, language (`--lang es/en/both`), and task weights by modifying the script.

The generation script is included in the repository (`generate_pymini.py`). Output is converted to Parquet for storage and distribution on the Hub.

## Intended Use

This dataset is intended for fine-tuning small language models (under ~2 billion parameters) to improve their Python understanding and code generation abilities. It is particularly suitable for:

- Code assistants that must explain or fix Python code.
- Educational tools that teach Python basics.
- Lightweight models running in resource-constrained environments (edge, mobile).

## Out-of-Scope Use

PyMini is not designed for:

- Large-scale production code generation.
- Mastering advanced Python libraries (NumPy, pandas, Django, etc.).
- Security-sensitive code auditing.

The examples are deliberately simple and may not cover edge cases of real-world software.

## Bias, Risks, and Limitations

- **Synthetic nature:** The dataset was generated from templates, so it may lack the stylistic variance of human-written code.
- **Language coverage:** While bilingual, the vocabulary is limited to common Python terminology. Colloquial expressions or complex natural language instructions may be underrepresented.
- **No malicious code:** The dataset does not include security vulnerabilities or harmful patterns; it is purely educational.
- **Potential overfitting:** Because of the templated generation, models may memorize patterns rather than generalizing to unseen code. Evaluate on diverse, real-world test sets.

## License

This dataset is released under the MIT License. You are free to use, modify, and distribute it for both research and commercial purposes.

## Citation

If you use PyMini in your work, please cite it as:

```bibtex
@misc{pymini2025,
  title        = {PyMini: A Compact Python Instruction Dataset for Small Language Models},
  author       = {Inserloft},
  year         = 2025,
  howpublished = {\url{https://huggingface.co/datasets/inserloft/PyMini}}
}
```

## Contributing

Feel free to open issues or pull requests if you have suggestions for new task types, additional languages, or improvements to the generation script.

## Contact

For questions, reach out via the Hugging Face community tab or through [inserloft.com](https://inserloft.com).