Revert to 7-label: F1 0.72 (binary was 0.28)

Binary classification: KEEP vs DISCARD (2 labels)

Phase 2b: Revert to Phase 1 + gentle label_smoothing=0.05

Phase 2: LR=1e-5, epochs=2, label_smoothing=0.1, warmup=0.15

Phase 1: max_length=768, batch=8, epochs=3, standard Trainer (no class weights)

Fix: Use fixthemusic namespace for model upload (token has access)

Fix: Use WeightedTrainer, calculate class weights, add early stopping, use bf16

Fix: Remove unsupported every parameter from Gradio

Upload app.py with huggingface_hub

Upload Dockerfile with huggingface_hub

Fix: Remove data/ copy since loading from HuggingFace Dataset

Upload README.md with huggingface_hub

Upload Dockerfile with huggingface_hub

Upload train.py with huggingface_hub

initial commit