OncoDIF
Model adaptation and evaluation for genomic curation and natural-language SQL
What I built
OncoDIF combines genomic evidence curation with natural-language clinical querying. I fine-tuned language models for this task and evaluated their outputs against defined baselines.
Post-training and evaluation
A Llama-3 model was fine-tuned with LoRA on 21,000 synthetic question–SQL pairs. On unseen questions, the study reported 83.24% correct SQL, compared with 55% for the untuned baseline.
The framework also evaluated terminology variation in simulated scenarios and flagged low-confidence outputs for review. These are task-specific results reported in a conference abstract; they do not establish clinical outcome improvements.
Publication
ASH 2025 conference abstract, published in Blood, volume 146, Supplement 1, abstract 2646.
