Fix total_mem -> total_memory attribute name

Robust v9: unbuffered output, error handling, no trackio/amp issues

Fix: train in fp32 to avoid T4 amp/bf16 issues

Fix bf16 LoRA params - cast after trainer creation

Fix bf16 error on T4 - cast all params to fp16

Add T4-optimized training script

Add comprehensive README with training details, usage, and research background

Add training script for financial options expert model

initial commit