Joesh1/onca-2.0-12b
Joesh1/onca-2.0-12b is a 12 billion parameter open oncology language model built on Google's Gemma 4 architecture, featuring a 262,144 token context window. It is specifically fine-tuned for oncology research workflows, particularly pancreatic cancer, excelling in trial screening, clinical reasoning, and variant evidence interpretation. This model is designed for specialized tasks like pathology extraction and provides improved performance over its predecessor, ONCA 1.5.
Loading preview...
ONCA 2.0: Specialized Oncology Language Model
ONCA 2.0 is an open oncology language model developed by Joesh1, based on google/gemma-4-12B-it. This 12 billion parameter model is specifically fine-tuned using a provenance-labeled oncology corpus, with a strong focus on pancreatic cancer and oncology-adjacent research workflows. It boasts an impressive 262,144 token context window, making it suitable for processing extensive clinical data.
Key Capabilities
- Trial Screening: Achieves 0.8240 accuracy in identifying eligible patients for clinical trials.
- Clinical Reasoning: Demonstrates 0.6761 outcome-label accuracy in complex clinical scenarios.
- Variant Evidence Interpretation: Provides 0.5427 macro-F1 for clinical significance.
- Pathology Extraction: While improved, this remains the model's primary weakness, with an overall field exact match of 0.4634.
When to Use
ONCA 2.0 is ideal for researchers and developers working with oncology data, particularly in the context of pancreatic cancer. It performs best with tightly scoped prompts and explicit output formats, making it suitable for:
- Automated trial screening.
- Assisting with clinical reasoning tasks.
- Extracting specific information from pathology reports.
- Interpreting variant evidence in a research setting.
It is important to note that this is a research model and its outputs require review by qualified experts; it is not intended as a clinical decision system.