"M3i team on AMD Developer Hackathon Hackathon"

Team Idea

Targeted LoRA improves accuracy on a held-out benchmark crypto-quant problems by a measurable margin vs base model AND vs random-layer LoRA control.

Team Members

suvasis mukherjee

suvasism

CEO

Minakshi Mukherjee

adaboost_mm

Submission

M3I: Where Does a Small LLM Fail, Layer-Wise?

Two LLMs audit a small MoE model's reasoning, identify failure-correlated layers via statistical attribution, and train LoRA-DPO targeting only those layers. Validated end-to-end on DeepSeek-V2-Lite; held-out eval pending controls.