MedReason-Stenographic

# MedReason-Stenographic: Medical QA with Compressed Reasoning Traces This dataset contains **31,535** medical question-answer pairs with stenographic reasoning traces, generated using MiniMax M2.1 from the original [UCSC-VLAA/MedReason](https://huggingface.co/datasets/UCSC-VLAA/MedReason) dataset. ## Dataset Description The dataset transforms medical QA reasoning into a **stenographic format** using a symbolic protocol designed for high-density, machine-parseable reasoning traces. This format eliminates natural language filler while maintaining the complete logical flow of medical reasoning. ### Key Features - **31,535 samples** of medical QA with compressed reasoning (filtered for non-empty reasoning) - **Stenographic reasoning traces** using symbolic notation - **Full thinking process** preserved in the `thinking` field - Based on the award-winning MedReason dataset (3rd place, HuggingFace Reasoning Datasets Competition 2025) ## Dataset Structure ### Fields | Field | Type | Description | |-------|------|-------------| | `query` | string | The medical question | | `reasoning` | string | Stenographic reasoning trace using symbolic notation (includes answer options inline) | | `answer` | string | The final answer in unified format (e.g., "D. Urea breath test") | | `thinking` | string | Full chain-of-thought reasoning before compression | | `id_in_dataset` | int | Reference ID from the original MedReason dataset | ### Example ```json { "query": "Most sensitive test for H pylori", "reasoning": "Detect(Task:RAG) ● → Domain(Gastroenterology) ● → Pathogen(H.pylori) ● → DiagnosticTests(SensitivityComparison) ● → UreaBreathTest(Sensitivity:95-98%) ● → FecalAntigen(Sensitivity:90-95%) ● → BiopsyUrease(Sensitivity:90-95%) ● → Serology(Sensitivity:80-90%) ● → DistractorAnalysis ※ → ComparativeAssessment ◐ → NonInvasiveGoldStandard ● → <H≈0.4> → ∴", "answer": "D. Urea breath test", "thinking": "The user is asking a medical question about diagnostic testing for Helicobacter pylori...", "id_in_dataset": 7134 } ``` ## Stenographic Protocol The reasoning traces use the **Grand Unified Reasoning Protocol (v4.1)** with the following symbolic lexicon: | Symbol | Meaning | Usage | |--------|---------|-------| | `→` | Flow/Derivation | Linear progression between reasoning steps | | `↺` | Refinement Loop | Self-correction or revisiting prior steps | | `∴` | Convergence | Final logical conclusion point | | `●` | Ground Truth | Verified fact from reliable source | | `◐` | Inference | Reasoned deduction not directly stated | | `○` | Speculation | Low-confidence hypothesis | | `※` | Constraint/Trap | Critical rule or limitation detected | | `⚠` | Risk/Warning | Hallucination risk or safety concern | | `<H≈X.X>` | Entropy Marker | Cognitive shift indicator (0.1=analytical, 1.5=creative) | ### Symbol Usage Statistics | Symbol | Samples | Percentage | |--------|---------|------------| | `→` | 31,419 | 99.5% | | `∴` | 30,996 | 98.2% | | `●` | 31,153 | 98.7% | | `<H≈` | 27,634 | 87.5% | | `※` | 18,045 | 57.2% | | `◐` | 16,777 | 53.1% | | `⚠` | 15,972 | 50.6% | | `↺` | 9,207 | 29.2% | | `○` | 1,622 | 5.1% | ## Dataset Statistics | Metric | Value | |--------|-------| | Total samples | 31,535 | | Unique source IDs | 10,550 | | Avg query length | 157.4 chars | | Avg reasoning length | 384.2 chars | | Avg answer length | 35.2 chars | | Avg thinking length | 5,548.4 chars | ### Data Quality - Empty queries: 89 (0.3%) - Empty reasoning: 0 (filtered out) - Empty answers: 9 (<0.1%) - All samples have thinking field populated ## Generation Process This dataset was generated using **MiniMax M2.1** with a specialized prompt that: 1. Takes the original MedReason question-answer pairs 2. Generates full chain-of-thought reasoning (`thinking` field) 3. Compresses the reasoning into stenographic format (`reasoning` field) 4. Outputs structured JSON with query, reasoning, and answer ## Comparison with Source Dataset | Aspect | Original MedReason | This Dataset | |--------|-------------------|--------------| | Size | 32,682 pairs | 31,535 pairs | | Reasoning format | Natural language | Stenographic symbols | | Thinking field | No | Yes (full CoT) | | Answer format | Text only | Unified "Letter. Text" format | | Fields | question, answer, reasoning, options | query, reasoning, answer, thinking, id_in_dataset | ## Use Cases 1. **Training reasoning models** with compressed, symbolic reasoning 2. **Fine-tuning for medical QA** with structured thought processes 3. **Research on reasoning compression** and symbolic representations 4. **Benchmarking** reasoning trace quality and consistency ## Citation If you use this dataset, please cite: ```bibtex @dataset{medreason_stenographic_2025, author = {Maziyar Panahi}, title = {MedReason-Stenographic: Medical QA with Compressed Reasoning Traces}, year = {2025}, publisher = {HuggingFace}, howpublished = {\url{https://huggingface.co/datasets/MaziyarPanahi/MedReason-Stenographic}} } ``` Also cite the original MedReason dataset: ```bibtex @dataset{medreason_2025, author = {UCSC-VLAA}, title = {MedReason}, year = {2025}, publisher = {HuggingFace}, howpublished = {\url{https://huggingface.co/datasets/UCSC-VLAA/MedReason}} } ``` ## License Apache 2.0