OTS Medical Training Data
MedHarm-Safety™
Physician reviewed
FDA anchored
Evidence-based-medicine
Standard-of-care
Built for medical AI labs, foundation-model teams, and pharmaceutical AI groups that require provenance—not scraped forums.
Benchmarks
17% → 38%
SFT
38% → 90%
Refined-tune
MedGemma 27B on MEDHARM-Bench
After training on MedHarm-Safety AE v1 the model named the event the FDA label actually warns about.
18% → 36%
SFT
36% → 82%
Refined-tune
Llama 3.3 70B on MEDHARM-Bench
After training on MedHarm-Safety AE v1 the model named the event the FDA label actually warns about.
SKU: MH-SAF-AE-CORE-v1-R1
MedHarm-Safety Adverse Effects v1
FDA-grounded adverse-event vignettes
A clean, labeled set that teaches models to name the *labeled* harm, not just the drug. FDA-anchored +updates. No PHI. Ready for LoRA or full SFT.
Model
SFT first pass
Refined-Tune second pass
MedGemma 27
17% → 38%
38% → 90%
Llama 3.3 70B
18% → 36%
36% → 82%
About
A clean physician-reviewed synthetic-hybrid dataset, anchored to evidence-based medicine, standard-of-care, and FDA-labeling (with scheduled updates) for training medical harm.
Harm categories
Pharmacotherapeutics, interventional, diagnostic, counseling, communication, preventive, ethical, and other high-value subsets.
Seeds
Seeds contain high-severity events and the clinical language that models actually need at training time.
About
Harm Taxonomy
- Adverse effects
- Contraindications
- Drug-drug interactions
- Dosing errors
- Duplication or deprescribing errors
- Controlled substances
- Monitoring errors
- Administration errors
- Wrong-site or wrong-procedure
- Planned
- Planned
- Planned
- Planned
Anchoring
All drugs are anchored to the FDA label website with regular updates available.
Request access.
The full data products catalogue are licensed, not listed for open download. Let us know who you are and drop us a message.