CRS draft adapter
LoRA adapter for Qwen/Qwen2.5-0.5B-Instruct, made for an individual university Generative AI assignment.
Educational screening prototype. Not a clinical enrollment decision. Do not use it for patient care.
- Task: given a numbered patient note, a criterion type and a criterion, return JSON with an eligibility label and supporting sentence IDs.
- Data: training split of TrialGPT Criterion Annotations, 602 rows from 37 patients. GPT-4 outputs in that dataset were never used.
- Settings: rank 32, alpha 64, oversampling power 0.5, label first output, best epoch 2.
- Validation (8 patients, 206 rows): accuracy 0.7379, macro-F1 0.4461, evidence micro-F1 0.4910. Held-out test results are in the repository.
- Code, split and usage: https://github.com/Masooma-Ali/PatientsDataScreener