Model Card Template
Example: Clinical Note Summarization v2.1
Model Details
| Field | Value |
|---|---|
| Model Name | clinical-summary-v2.1 |
| Base Model | Claude 3.5 Sonnet (claude-3-5-sonnet-20241022) |
| Fine-tuned | No (prompt-engineered) |
| Version | 2.1.0 |
| Release Date | 2025-10-15 |
| Owner | AI Platform Team |
| Contact | ai-platform@company.com |
Intended Use
Primary use case: Summarizing clinical notes for physician review dashboard.
Intended users: Clinical staff reviewing patient histories.
Out of scope: - Direct patient communication - Clinical decision making without physician review - Pediatric notes (not validated) - Non-English notes
Training/Prompt Data
- Prompt developed using 500 de-identified clinical notes
- Validated on 200 held-out notes with physician review
- No fine-tuning performed; uses prompt engineering only
Performance Metrics
| Metric | Value | Measurement Date |
|---|---|---|
| Physician approval rate | 94.2% | 2025-10-01 |
| Factual accuracy (spot check) | 98.1% | 2025-10-01 |
| Hallucination rate | 1.2% | 2025-10-01 |
| Mean processing time | 2.3s | 2025-10-01 |
| P95 latency | 4.1s | 2025-10-01 |
Limitations
- May miss nuanced clinical context requiring specialist knowledge
- Abbreviation expansion occasionally incorrect for rare terms
- Long notes (>10 pages) may have reduced summary quality
- Not validated for: surgical notes, radiology reports, pathology
Ethical Considerations
- All outputs must be reviewed by licensed clinician
- PHI handling compliant with HIPAA
- Bias testing performed across demographic groups (see appendix)
- No disparate impact detected in summary quality by patient demographics
Monitoring
- Daily evaluation on 100 production samples
- Weekly physician review of flagged cases
- Monthly bias audit
- Alerts: >5% drop in approval rate, >3% hallucination rate
Version History
| Version | Date | Changes |
|---|---|---|
| 2.1.0 | 2025-10-15 | Improved medication list extraction |
| 2.0.0 | 2025-08-01 | Switched to Claude 3.5, new prompt |
| 1.2.0 | 2025-05-01 | Added allergy highlighting |
Model Card Template (Blank)
# Model Card: [Model Name] {.unnumbered}
## Model Details
| Field | Value |
|-------|-------|
| Model Name | |
| Base Model | |
| Version | |
| Release Date | |
| Owner | |
| Contact | |
## Intended Use
**Primary use case:**
**Intended users:**
**Out of scope:**
## Training/Prompt Data
[Description of data used for training or prompt development]
## Performance Metrics
| Metric | Value | Measurement Date |
|--------|-------|------------------|
| | | |
## Limitations
[Known limitations and edge cases]
## Ethical Considerations
[Bias testing, fairness considerations, required human oversight]
## Monitoring
[Ongoing evaluation cadence, alert thresholds]
## Version History
| Version | Date | Changes |
|---------|------|---------|
| | | |Model Card Best Practices
What to include: - Specific, measurable metrics with dates - Explicit scope boundaries (what it’s NOT for) - Known failure modes - Required human oversight - Monitoring and alerting setup
What to avoid: - Vague claims (“high accuracy”) - Missing limitations - No versioning - Static documents that aren’t updated