mkurman/medical-reasoning-synthlabs-I

Medical Reasoning SynthLabs I

Medical Reasoning SynthLabs I is a synthetic medical reasoning dataset generated with SynthLabs. It contains medical question-answer examples paired with structured reasoning traces, intended for research on reasoning-style instruction tuning, medical QA, answer synthesis, and reasoning trace analysis.

The dataset is designed for machine learning research and experimentation. It is not intended for clinical decision-making, diagnosis, treatment planning, or use as a source of medical advice.

Dataset Details

• —Repository: mkurman/medical-reasoning-synthlabs-I
• —Rows: 856 examples
• —Format: Parquet
• —Split: train
• —Modality: Text
• —Generation model label: DEEP: qwen3-synth
• —Primary task type: Medical question answering with reasoning

Dataset Structure

The dataset contains one split:

Columns

Example Usage

from datasets import load_dataset

dataset = load_dataset("mkurman/medical-reasoning-synthlabs-I")

print(dataset)
print(dataset["train"][0])

Intended Uses

This dataset may be useful for:

• —Fine-tuning language models on structured medical reasoning.
• —Training models to produce concise answers from longer reasoning traces.
• —Evaluating reasoning-trace formatting and answer consistency.
• —Building synthetic-data pipelines for domain-specific QA.
• —Researching medical instruction tuning and reasoning supervision.

Out-of-Scope Uses

This dataset should not be used for:

• —Direct clinical diagnosis or treatment recommendations.
• —Replacing medical professionals or validated clinical guidelines.
• —Patient-facing medical advice without expert review.
• —High-stakes medical deployment without rigorous validation.
• —Benchmarking real-world clinical safety without additional expert evaluation.

Data Generation

The examples were generated using SynthLabs in a medical reasoning format. Each sample contains a medical query, a structured reasoning trace, and a final answer.

The visible generation label in the dataset is:

DEEP: qwen3-synth

This dataset was created through a two-stage synthetic generation process.

First, medical topics were provided to moonshotai/kimi-k2, which generated long-form medical text about those topics. These generated texts were then used as source material for a second synthesis step.

In the second step, qwen/Qwen3-14B was used to generate SynthLabs-style reasoning and answer pairs based on the long-form medical texts. Each final example contains a medical query, a reasoning trace, and a final answer.

Finally, the generated examples were evaluated and filtered using deepseek-ai/DeepSeek-V3 to improve overall quality and remove lower-quality generations.

This means the dataset should be understood as synthetic medical reasoning data derived from model-generated medical topic expansions, rather than from human-authored clinical cases or real patient records.

⚠️ Because the source medical text, reasoning traces, and answer pairs were generated through language-model-based synthesis, the dataset may still contain inaccuracies even after final evaluation and filtering. The deepseek-ai/DeepSeek-V3 filtering step is intended to improve quality, but it should not be treated as clinical validation.

Limitations

Because this is a synthetic dataset, it may contain:

• —Medical inaccuracies.
• —Hallucinated facts or unsupported assumptions.
• —Overconfident reasoning.
• —Incomplete differential diagnoses.
• —Reasoning traces that appear plausible but are not clinically validated.
• —Biases inherited from the source prompts, generation model, or synthesis pipeline.

Users should perform medical expert review before using this dataset for any health-related application.

Ethical Considerations

Medical reasoning datasets require extra caution because incorrect outputs can cause harm if used in real-world healthcare settings. This dataset is intended for research, prototyping, and model development only.

Recommended safeguards include:

• —Expert clinical review.
• —Automated and manual quality checks.
• —Filtering for unsafe or misleading generations.
• —Clear downstream disclaimers.
• —Evaluation against trusted medical references and guidelines.

Citation

If you use this dataset, please cite it as:

@misc{kurman2026medical_reasoning_synthlabs_i,
  title        = {Medical Reasoning SynthLabs I},
  author       = {Mariusz Kurman},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/mkurman/medical-reasoning-synthlabs-I}}
}

License

This dataset is released under the Apache 2.0 license.

Disclaimer

This dataset is for research and educational use only. It is not medical advice and should not be used for diagnosis, treatment, triage, or other clinical decision-making.