Full training results & evaluation with BERTScore

Fixed compiling issue, added legnth penalty, and atttempting freezing encoder layers 0-5 to lower parameters and preserve T5's langauge understanding.

Clean up codebase and fix training bugs

Update LexiMind: improved training, model architecture, and evaluation

feat: Add FLAN-T5 compatibility with relative position bias

chore: snapshot current refinements

Reformatted Project Structure