Update README with corrected claims, add competition writeup

Revert: remove eval from startup, restore normal operation

Add LoRA evaluation on startup (RUN_LORA_EVAL flag)

Revert: remove training from startup, keep torch env fixes

Fix: remove unsupported SFTConfig params

Fix: USER env for torch cache, remove duplicate import

Fix: use SFTConfig for newer TRL API

Fix: remove deprecated dataset_text_field for newer TRL

Fix: tokenizer -> processing_class for newer TRL

Fix: set USER env var for PyTorch cache in container

Fix: total_mem -> total_memory in training script

Add optional LoRA training on startup (RUN_LORA_TRAINING flag)

Restore start.sh script