MODEL ENGINEERING
MedTech Genomic Distillation
MedTech Genomic Distillation is an end-to-end model engineering research prototype. The pipeline ingests ClinVar variant data, generates reasoning traces with a DeepSeek R1 teacher, keeps only the traces a verifier agrees with, and trains Qwen3 8B and 4B students with LoRA on Apple Silicon through MLX. Classical ML baselines and an evaluation pipeline sit next to the students so the results can be compared honestly. The 8B student reaches 89.80% on the 98-case validation split and 88.89% on a separate 45-case holdout that was never distilled. This is a research prototype, not a clinically validated system.
What It Demonstrates
ClinVar ingestion with verifier-guided filtering of teacher reasoning traces
Qwen3 8B and 4B LoRA training through MLX on Apple Silicon
Classical ML baselines scored against the same splits as the students
8B student at 89.80% validation on 98 cases and 88.89% on a 45-case never-distilled holdout
A research prototype with an explicit evaluation pipeline, not a clinically validated system
Results
