| --- |
| license: agpl-3.0 |
| license_name: sspl-1.0 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| - split: validation |
| path: data/validation-* |
| dataset_info: |
| features: |
| - name: messages |
| list: |
| - name: role |
| dtype: string |
| - name: content |
| dtype: string |
| - name: metadata |
| struct: |
| - name: topic_path |
| list: string |
| - name: checker_score |
| dtype: int64 |
| splits: |
| - name: train |
| num_bytes: 135701582 |
| num_examples: 18098 |
| - name: validation |
| num_bytes: 16079590 |
| num_examples: 2012 |
| download_size: 151672373 |
| dataset_size: 151781172 |
| --- |
| |
| # pythink-20k |
|
|
| **High-quality Python reasoning traces for training rigorous, step-by-step thinking.** |
|
|
| `pythink-20k` is a quality-controlled, synthetic, curated dataset of **20,000 executable Python reasoning plans**. Each example teaches a model to |
| break down complex problems into structured, verifiable steps using a consistent execution trace style. |
|
|
| This is an initial release focused on getting an initial dataset out for testing and experimentation. I expect that the quality is pretty good, but |
| quality is currently limited by the ability for existing models to represent thoughts as Python code. |
|
|
| ## Why? |
|
|
| Current LLMs have very unstructured, undisciplined thoughts which are poorly comprehensible and often inefficient. Using a programming |
| language for thinking makes thoughts easier to audit, more structured, and more consistent. |
|
|
| Why Python? |
| - LLMs are excellent at Python already |
| - Python is more token-efficient than JSON, typescript, or other languages |
|
|
| ## What’s in the Data? |
|
|
| Each example contains a full Python-based reasoning trace that: |
|
|
| - Uses short, action-oriented `# Next:` comments |
| - Performs actual modeling, calculation, validation, or simulation in code |
| - Stays in "plan mode" (does not output raw implementation code in other languages) |
| - Ends with a clean `finalAnswer` |
|
|
| The training data treats the **entire reasoning trace** as its response. This dataset will NOT train a model to answer the question. The whole |
| point is to train the model to consider **how the question should be thought about**, instead of rushing to reach a conclusion. |
|
|
| The outputs are generally executable Python code, meaning than many of the responses can be executed to get the real output. |
|
|
| ### Example Format |
|
|
| ```json |
| { |
| "messages": [ |
| { |
| "role": "user", |
| "content": "Design a system that programmatically generates a TLS 2048-bit RSA certificate with 90-day validity..." |
| }, |
| { |
| "role": "assistant", |
| "content": "# Next: validate RSA certificate requirements...\nrsaKeyBits = 2048\nvalidityDays = 90\n..." |
| } |
| ], |
| "metadata": { |
| "topic_path": ["coding", "Network Servers and Daemons", "TLS/SSL Termination"], |
| "checker_score": 10 |
| } |
| } |
| ``` |
|
|
| ## Key Features |
|
|
| - **20k high-quality traces** (filtered with automated quality gate) |
| - Strong emphasis on **executable reasoning** rather than descriptive text |
| - Diverse domains: mathematics, coding, physics, philosophy, biology, and more |
| - Includes rich metadata (`checker_score`, topic path) |
| - Clean chat format ready for SFT |
|
|
| ## Intended Use |
|
|
| This dataset is designed for: |
|
|
| - Supervised fine-tuning of reasoning models |
| - Training models to produce structured, verifiable reasoning traces |
| - Research on executable chain-of-thought and code-as-reasoning |
| - Building the foundation for more advanced agentic and memory-augmented systems |
|
|
| ## Data Quality |
|
|
| All examples passed a multi-criteria quality filter including: |
| - Minimum checker score |
| - Presence of executable steps and validation logic |
| - Avoidance of target-language code leakage |
| - Reasonable length and structure |
|
|
| ## Limitations |
|
|
| - Some traces may still contain minor stylistic inconsistencies. |
| - Focused on reasoning structure rather than factual correctness. |
|
|
| ## License |
|
|
| AGPL-3.0 |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite: |
|
|
| ```bibtex |
| @misc{pythink-20k, |
| author = {NuclearManD}, |
| title = {pythink-20k: High-Quality Python Reasoning Traces}, |
| year = {2026}, |
| howpublished = {\url{https://huggingface.co/datasets/NuclearManD/pythink-20k}} |
| } |
| ``` |
|
|
| ## Future Releases |
|
|
| Planned improvements include: |
| - Higher quality filtering and human review |
| - Preference data (chosen vs rejected reasoning traces) |
| - Multi-turn and agentic reasoning examples |
|
|
| --- |
|
|
| **Built as part of ongoing work on rigorous reasoning systems.** |
|
|
