Strip leaked W&B API key from training notebook

Sync chunk: training scripts + notebook

Sync latest submission updates (GRPO-only + blog markdown + bugfixes)

Improve training data quality, teacher policy, ablation, and reward-aligned stage

Add LoRA-enabled script runner with auth-safe snapshot download

Add eval progress prints for long runs

Improve notebook runner logging and timeout safety

Use LoRA adapter tuning to avoid T4 OOM

Sync latest notebook training fixes

Add notebook runner script for HF Jobs

Use pure fp16 weights without AMP scaler in SFT

Disable grad clipping for stable fp16 SFT on T4

Fix CUDA dtype and device handling for full training

Enable GPU-friendly full training config

Upload folder using huggingface_hub