hetanshwaghela/autoscientist-healthcare-reasoning

🩺 Adapted Healthcare Clinical-Reasoning (AutoScientist)

Built with Adaptive Data by Adaption. A grounded, safety-blueprinted clinical-reasoning dataset — and a rigorous, fully-reproducible study of when data adaptation helps a small model, and when it doesn't.

<p align="center">

</p>

20,000 rows of English reason-then-answer clinical reasoning across specialties (cardiology, oncology, endocrinology, infectious disease, psychiatry, and more), grounded under an explicit clinical-safety blueprint and blended with curated public-health operations examples.

✨ Why this dataset

Most "medical LLM" datasets optimize for a final answer. This one optimizes for the reasoning that earns the answer — and for doing it safely: grounded in evidence, hedged appropriately, escalating red flags, and refusing when uncertain. Every row is prompt → step-by-step reasoning → justified conclusion, then hardened with Adaptive Data's hallucination_mitigation under a clinical blueprint.

The result is a +30% platform-measured quality lift (Grade B→A) with the largest gain exactly where it matters clinically — completion quality +37.9%.

🔬 Provenance pipeline (reproducible end-to-end)

FreedomIntelligence/medical-o1-reasoning-SFT   (Apache-2.0, gold chain-of-thought)
        │   prompt = Question
        │   completion = Complex_CoT + "\n\nFinal answer:\n" + Response
        ▼
+ 30 curated CDC-grounded public-health operations rows   (safe escalation / refusal / grounding)
        ▼
Adaptive Data by Adaption   →   deduplication + hallucination_mitigation
        │   steered by a clinical-reasoning safety blueprint
        ▼
Adapted dataset (20,000 rows)   ──►   AutoScientist training (Qwen3.5-0.8B, LoRA)

• —Dataset id: 26048b57-f164-46d5-810b-12d498a76660
• —Clinical safety blueprint: ground in evidence; never fabricate stats / dosages / citations; preserve numbers exactly; present as educational reasoning, not diagnosis; recommend clinician confirmation; flag emergencies; state uncertainty.

🧬 Data fields

📦 Splits

• —train — 20,000 rows (avg completion ≈ 780 words).
• —A small held-out private evaluation set (public-health operations, source- and scenario-disjoint) is kept off-platform and not released, to prevent leakage.

🧪 The honest science: what we learned training on it

We trained the companion Qwen/Qwen3.5-0.8B (LoRA) on this data and evaluated it head-to-head against its base with an LLM judge (Gemini 3.1 pro, 200 held-out samples). The base model won (58/42). We didn't stop there — we ran a second, larger run (60k rows, +40% general-purpose data, tuned recipe) and it landed lower (62/38).

Two clean runs, one conclusion — a reproducible finding, not a fluke:

*Supervised fine-tuning a 0.8B already-instruction-tuned model on long (≈780-word) chain-of-thought increases verbosity, and the judge prefers the base's crisper answers. Dataset quality (+30%) and small-model head-to-head win-rate are different axes — a better dataset does not automatically beat a strong small base.*

Evidence (see the model card): eval-loss plateaued after ~epoch 1 while train-loss kept falling (overfitting), and a higher-LR / more-steps recipe moved further from base and lost by more.

This is exactly the kind of result the AutoScientist framing is built to surface: the dataset is the contribution; the model is the honestly-reported experiment.

✅ Intended use

• —Fine-tuning / evaluating assistants for grounded clinical reasoning and public-health operations guidance.
• —Research on chain-of-thought, hallucination mitigation, grounding, refusal, and the dataset-quality-vs-small-model-capability question above.

⚠️ Out-of-scope, limitations & safety

• —Educational decision-support, not a medical device. Not for individual diagnosis, treatment, or medication dosing without a qualified clinician.
• —Reasoning is partly machine-generated (GPT-4o + medical verifier upstream, then Adaptive Data); residual errors possible. English, exam-style skew.
• —Guidance can become superseded; defer to current official sources.
• —Present conclusions as reasoning to be confirmed; never fabricate stats/dosages/ citations; preserve numeric values exactly; escalate emergencies.

📜 Source, attribution & licensing

• —Foundation: FreedomIntelligence/medical-o1-reasoning-SFT — Apache-2.0 (redistribution permitted with attribution). This adapted derivative: Apache-2.0.
• —Public-health blend: U.S. federal public-domain text. Source: Centers for Disease Control and Prevention. No CDC endorsement is implied.
• —Adapted with Adaptive Data by Adaption.

🔗 Companion artifacts

• —🧪 Live demo (HF Space): https://huggingface.co/spaces/hetanshwaghela/autoscientist-healthcare-demo
• —Model (LoRA): https://huggingface.co/hetanshwaghela/autoscientist-healthcare-reasoning
• —Kaggle dataset: https://www.kaggle.com/datasets/hetanshwaghela1/autoscientist-healthcare-reasoning
• —Kaggle model: https://www.kaggle.com/models/hetanshwaghela1/autoscientist-healthcare-reasoning

Built with Adaptive Data by Adaption. Foundation: FreedomIntelligence/medical-o1-reasoning-SFT (Apache-2.0). Public-health blend: CDC public-domain text.