Add extend_dataset.py: YouTube scraping + ESC-50 + expanded synthetic (run locally for YouTube)

Update README: clean clone-and-run instructions

Update requirements.txt with all dependencies

Update run_ablations.py: uses new train.py with Hub dataset + OGM-GE

Overwrite train.py: defaults to Hub dataset (Ellaft/pc-fault-real-dataset), uses models.py + dataset_v2.py, OGM-GE enabled

Overwrite models.py with v2 architecture (auxiliary heads + OGM-GE + anti-collapse)

Overwrite config.py with v2 config (OGM-GE params, asymmetric LR, auxiliary loss weights)

Update train_v2.py: add --dataset flag to switch between old proxy data (dataset_real) and new built dataset (dataset_v2)

Add dataset_v2.py: adapter for build_dataset.py output, drop-in replacement for dataset_real.py

Add build_dataset.py: YouTube scraping + HF datasets + synthetic generation pipeline (1400+ lines, no API keys needed)

Add build_dataset.py: Complete data pipeline (YouTube scraping + HF datasets + synthetic generation)"

Add run_ablations_v2.py: v2 ablation runner with OGM-GE experiments

Add train_v2.py: OGM-GE training loop + asymmetric LRs + aux loss logging

Add models_v2.py: auxiliary heads + OGM-GE anti-collapse

Add config_v2.py: OGM-GE hyperparameters + asymmetric LR config

Add real dataset loader (ToyADMOS audio + MVTec AD visual)

Add flexible dataset module supporting custom HF audio and image datasets

Add requirements.txt

Add ablation runner script

Add training script and ablation runner

Add model architecture (ViT + AST + Late Fusion + LoRA)

Add dataset, models, training, and ablation scripts

Add configuration module

Add project README with architecture and experiment design

initial commit