chore: test trial run for colab script with 10 dataset size and fixed dependencies

fix: downgrade TRL to 0.15.0 and nuke ghost vLLM install to prevent import crashes

fix: correct trl version pin to 0.17.0

fix: vLLM/TRL version conflict in Colab Cell 1

fix: training reward bounds to (0.01, 0.99) — prevents loss=0 from zero variance

fix: increase num_generations=6, temperature=0.9, lr=2e-5 to fix zero-loss GRPO

fix: remove fp16=True, use paged_adamw_8bit for QLoRA BFloat16 compatibility

fix: add 4-bit quantization + LoRA to fit GRPO training in T4 15GB VRAM

refactor: rewrite training script - remove rollout_func, use inline reward evaluation

fix: remove deprecated GRPOConfig params for latest TRL compatibility

fix: set live HF Space URL in training script and demo app

feat: CodeSensei - GRPO-trained LLM code debugger on OpenEnv