rollback: revert to last working Dockerfile and train.py

fix: proper GRPO with trl 0.12.2 no-deps + force hub downgrade

fix: custom training loop without TRL dependency

fix: trl 0.12.2 has GRPOTrainer, pin all deps before trl install

fix: try multiple import paths for GRPOConfig

fix: trl 0.11.4 + transformers 4.46.0 + processing_class

fix: trl 0.9.4 + transformers 4.41.2 compatible versions

fix: remove tokenizer arg from GRPOTrainer

fix: tokenizer not processing_class, torch cu121 for GPU

fix: CPU fallback when no GPU detected

improve: abstention penalty, better prompt, mixed curriculum, more steps

update: 500 steps L1 + 300 steps L2, higher lr for 1.5B

fix: remove misplaced import inside GRPOConfig args

fix: auto-detect bf16 support

fix: copy data to multiple locations, fallback path for level2

fix: use dynamic path for level2.jsonl

fix: use /tmp for all output directories

fix: patch getpwuid error for HF Spaces

add: Level 2 training for 1.5B model after Level 1

fix: batch size 4 to match num_generations 4

fix: replace unsloth with standard transformers+peft, no version conflicts

fix: install deps in Dockerfile build, not runtime

fix: use /tmp for cache dirs, run train.py

upgrade: Qwen 1.5B model, 150 L1 + 80 L2 steps

Add health server on port 7860 for HF Spaces keep-alive

update: 200 steps L1 + 100 steps L2 training

fix: download datasets from GitHub at runtime instead of relying on package data

fix: install unsloth_zoo and nest-asyncio properly

feat: DECEIT Level 1 GRPO training job