Joesh1/onca-3.0-12b

ONCA 3.0-12B: Specialized Oncology Research Model

ONCA 3.0-12B is an 11.95 billion parameter research model, fine-tuned from google/gemma-4-12B, dedicated to structured oncology research tasks. It processes text-only inputs and generates compact JSON outputs, streamlining the analysis of complex medical data.

Key Capabilities

• Pathological Classification: Accurately performs TNM (Tumor, Node, Metastasis) classification for pathological reports.
• Clinical Trial Support: Handles clinical-trial triage and criterion reasoning, aiding in patient selection and study design.
• Molecular Evidence Classification: Classifies CIViC molecular-evidence, contributing to precision oncology efforts.
• Structured Output: Designed to produce consistent, schema-valid JSON outputs for all six supported tasks, facilitating integration into automated workflows.

Performance Highlights

The model demonstrates strong exact match performance across its six core tasks on a 4,599-row development-validation set, achieving an overall exact match of 85.02% and a schema validity of 99.83%. For instance, it reached 95.06% exact match for Pathological M classification and 92.43% for Pathological N classification.

Training and Limitations

ONCA 3.0-12B underwent two phases of NF4 QLoRA SFT on 16,625 unique examples, followed by a GRPO stage. It's important to note that this is a research model and not a medical device; human oncology/pathology validation is still pending. The model can hallucinate or misclassify, and results are based on development data, not locked final test partitions. It is text-only and should not be used with images or identifiable patient information.