MAAS GRPO Run

• model: Qwen/Qwen2.5-0.5B-Instruct
• epochs: 1
• batch size: 6
• gradient accumulation steps: 1
• num generations: 6
• learning rate: 5e-06
• temperature: 1.0
• top_p: 0.95
• dataset size: 21
• prompt format: chat_messages
• log completions: True
• output dir: artifacts/maas-grpo-hackathon-final
• train metrics: {"total_flos": 0.0, "train_loss": 0.04405292901986589, "train_runtime": 82.4663, "train_samples_per_second": 2.183, "train_steps_per_second": 0.364}