Gemma3-270M-EU-Medical-CandidateSelector-LoRA

LoRA adapter for unsloth/gemma-3-270m-it.

This model is a candidate-conditioned German/EU medical terminology selector.

It is not a diagnostic model. It must not be used for treatment, prescribing, dosage, triage, or clinical decision-making.

Intended role

Verifier retrieves ICD/EMA candidates → 270M LoRA selects or flags ambiguity → Verifier confirms selected candidate → MedGemma explains verified facts.

Training

• Base model: unsloth/gemma-3-270m-it
• Dataset: MediNorm-DE-EU-TeacherSelect-100K
• Epochs: 2
• Steps: 2,624
• Effective batch size: 64
• Trainable parameters: 3,796,992 / 271,895,168
• Final validation loss: 0.001195

Full held-out test results

Evaluated on 8,311 rows:

• JSON parse rate: 0.9998
• Task exact rate: 1.0000
• Safety-scope exact rate: 1.0000
• ICD selected-code exact rate: 0.9956
• Selected-from-candidate-set rate: 1.0000
• Invalid out-of-candidate code rate: 0.0000
• ICD no-match null rate: 1.0000
• ICD ambiguity code exact rate: 0.9970
• ICD ambiguity terminal exact rate: 0.9985
• EMA medicine selection exact rate: 1.0000
• EMA active-substance exact rate: 0.9702
• EMA ATC exact rate: 0.9979
• Safety refusal rate: 1.0000

Trust boundary

Trusted for:

• valid JSON
• candidate selection from provided lists
• ambiguity/non-terminal flagging
• terminology-only safe outputs

Not trusted for:

• free-generation of ICD codes
• free-generation of EMA metadata
• diagnosis
• treatment advice
• dosage advice

Production rule

The selected code or medicine must be verified against deterministic SQLite/EMA tables and must be present in the provided candidate list.

The verifier should overwrite metadata fields from deterministic sources:

• title_de
• terminal
• chapter
• active_substance
• inn_or_common_name
• atc_code
• medicine_status