| # De-identification Benchmark Results | |
| **Model:** Minibase-DeId-Small | |
| **Dataset:** Personal_De-identifier_Benchmark_SFT.jsonl | |
| **Sample Size:** 100 | |
| **Date:** 2025-09-25T12:38:54.363196 | |
| ## Overall Performance | |
| | Metric | Score | Description | | |
| |--------|-------|-------------| | |
| | PII Detection Rate | 1.000 | How well personal identifiers are detected | | |
| | Completeness Score | 0.670 | Percentage of texts fully de-identified | | |
| | Semantic Preservation | 0.109 | How well meaning is preserved | | |
| | Average Latency | 483.7ms | Response time performance | | |
| ## Key Improvements | |
| - **PII Detection**: Now measures if model generates ANY placeholders when PII is present in input | |
| - **Unified Evaluation**: All examples evaluated together (no domain separation) | |
| - **Lenient Scoring**: Focuses on detection capability rather than exact placeholder matching | |
| ## Example Results | |
| ### Example 1 | |
| **Input:** Patient Sarah Johnson, DOB 05/12/1980, visited Dr. Lee at St. Jude Hospital on 2023-10-26. Her conta... | |
| **Expected:** Patient [NAME_1], DOB [DOB_1], visited [NAME_2] at [HOSPITAL_1] on [DATE_1]. Her contact is [PHONE_1... | |
| **Predicted:** Patient [FIRSTNAME_1] [MIDDLENAME_1], DOB [DOB_1], visited Dr. [LASTNAME_1] at [CITY_1] Hospital on ... | |
| **PII Detection:** 1.000 | |
| ### Example 2 | |
| **Input:** Deponent Mr. Robert Davis, CEO of GlobalCorp Inc., stated under oath on December 1, 2022, that his a... | |
| **Expected:** Deponent [NAME_1], CEO of [ORGANIZATION_1], stated under oath on [DATE_1], that his attorney, [NAME_... | |
| **Predicted:** Deponent [PREFIX_1] [FIRSTNAME_1] [LASTNAME_1], CEO of [COMPANYNAME_1], stated under oath on [DATE_1... | |
| **PII Detection:** 1.000 | |
| ### Example 3 | |
| **Input:** Employee ID: EMP-001-XYZ. Name: John Doe. Salary: $85,000. Email: [email protected]. Marital Stat... | |
| **Expected:** Employee ID: [EMPLOYEE_ID_1]. Name: [NAME_1]. Salary: [SALARY_1]. Email: [EMAIL_1]. Marital Status: ... | |
| **Predicted:** Employee ID: EMP-[BUILDINGNUMBER_1]. Name: [FIRSTNAME_1] Doe. Salary: [CURRENCYSYMBOL_1][AMOUNT_1]. ... | |
| **PII Detection:** 1.000 | |