Polish for hackathon submission: training evidence, two pipelines, UI, docs

Fix health check timeout: start UI server in background before training

Reduce prompt/completion length to fix silent OOM on backward pass

Replace env-simulation reward with fast pure-heuristic to fix hang

Fix batch/num_generations: 4/4 on 1 GPU, grad_accum=4

Fix GRPO training: reward variance, batch/gen alignment, generation config

Update run_training.py and train_grpo.py, remove Dockerfile.training

Add pre-train gen sanity check, explicit GenerationConfig, dynamic GRPOConfig params, torch_compile/vllm off

QLoRA best practices: prepare_model_for_kbit_training, paged_adamw_8bit, cosine LR, faster iteration

Fix: enable_input_require_grads for gradient checkpointing + 4-bit

Fix OOM: reduce batch/gen/tokens, add grad checkpointing + adafactor

Fix batch size: 8 to match num_generations=8

Drop unsloth: use standard bitsandbytes 4-bit + peft LoRA + TRL GRPOTrainer

Add explicit compiler location debugging in run_training.py

GRPO training with CUDA + results in UI

feat: curriculum training + Karnataka scenarios + repo cleanup

fix: total_memory attribute name

Add GRPO training runner for HF Spaces GPU