Fix: GRPO not in trl 0.11, need trl>=0.14.0

Fix trl/pytorch version incompatibility + indentation bugs

Fix indentation bug on line 42

Fix indentation bug in grpo_train.py + update requirements.txt

HF Spaces GPU training pipeline

Update blog with 0.5B results and project metrics

Update blog with 0.5B results and project metrics

Update blog with 0.5B results and project metrics

Update blog with 0.5B results and project metrics

..

..

Revert Dockerfile to serve web app instead of training (per professor's advice)

Fix HF Space crashing: Copy missing training directory and fix user 1000 permissions

Fix HF Space crashing: Resolve port mismatch, fix CRLF line endings, and force Python unbuffered output

Automate 7B Training using Hugging Face Space Dockerfile

Fix FP16 AMP crash by explicitly loading base model in float32 for fallback hardware

Fix BFloat16 AMP crash by explicitly casting to float16 during fallback loading

Fix bf16 error for Colab T4 compatibility

Fix GRPOConfig __post_init__ crash by ensuring batch_size matches num_generations

..

feat: scale up to Qwen2.5-7B, set GRPO steps to 150 for health check, add HF push cell

fix: save reward_history.txt from GRPO trainer logs after --mode grpo

..

fix: add training dir to sys.path so -m training.test_rollout works on Colab

fix: colab working dir bug, rollout sys.path, openenv imports, add plot_rewards

first commit