VoxelZero

VoxelZero is an open research initiative for medical AI founded and maintained by Krish Malik, Senior @ Parkland High School, Allentown, PA
I create and publish curated datasets and AI frameworks for radiology and multimodal AI research.

🔗 Resources

• GitHub: github.com/krishm-12
• Featured Dataset: ROCO v2 Caption Subset (1k JSONL)
• VLM-ModalityFlip (framework): https://github.com/krishm-12/VLM-ModalityFlip

VLM-ModalityFlip

VLM-ModalityFlip is a research framework for probing modality dominance and representation shifts in Vision–Language Models (VLMs), with a focus on medical imaging tasks. The framework enables:

• Layer-wise analysis of visual vs textual signal dominance
• Identification of modality “flip” points across model depth
• Quantitative metrics and visualizations for interpretability
• Reproducible evaluation pipelines for VLM behavior and hallucination analysis

🔗 Full source code & documentation:

👉 https://github.com/krishm-12/VLM-ModalityFlip

📄 Research Note

VoxelZero tools are intended for:

• academic research
• benchmarking and interpretability studies
• medical AI methodology exploration