Back to work
CSE443: Bioinformatics · 2025
BioAlign-QLoRA: Biomedical Knowledge Graph Alignment
Curated 68,444+ gene-disease relationships and adapted Llama-, Mistral-, and Phi-style models with QLoRA to study changes in biomedical embedding alignment.
[ EVD ]
Engineering evidence
- My contribution
- Lead researcher · Data curation, QLoRA experiments, and evaluation
- Outcome
- Custom embedding-space separation metric and comparative model study
- Decision record 01
- Used parameter-efficient QLoRA to compare multiple model families within constrained research compute.
[ KEY ]
Technical highlights
- 01
Introduced a novel 'Knowledge Graph Separation' score quantifying geometric alignment between LLM embedding space and biological knowledge structures.
- 02
Achieved 83.8% accuracy with Mistral model, outperforming the pre-trained BioMistral-7B expert.
[ STK ]
Stack
- Python
- PyTorch
- QLoRA
- Llama-3
- Mistral-7B
- Knowledge Graphs