Refactor code structure for improved readability and maintainability and also added the mask ROI

Refactor A1 evaluation protocols to focus exclusively on Protocol C; update documentation and scripts accordingly.

Skip LoRA SFT for LLaDA (masked diffusion LM, incompatible with causal SFT); pass through HF ID

Fix TRL compat: pre-tokenize dataset, drop max_seq_length from SFTConfig/SFTTrainer

Fix SFTConfig/SFTTrainer API compat across TRL versions (max_seq_length, processing_class)

Fix duplicate --feature-num-workers/--num-fit-workers args in parser

Enhance feature extraction with multi-GPU support and worker resolution; add parallel processing for fitting layers

Add instruct-tuning (LoRA), multi-model fit, comparison viz; fix missing flags

added the test run for HF job

Add core ROI masks, visualization script, and model profiles

Merge hf/master: restore run_a1 files and hf_jobs