brosnanyuen/Qwen3.8-27B-GRPO-main-defender

Qwen3.8-27B GRPO main defender

Stage-5 GRPO (Unsloth + TRL) LoRA adapter for the Qwen3.8-27B cyber-range defender policy.

• Run: main round 1
• Base model: unsloth/Qwen3.8-27B-unsloth-bnb-4bit
• Warm start (SFT v3): brosnanyuen/Qwen3.8-27B-SFT-Defender-v3

metric	value
train_loss	0.0003113255661446601
reward	None
reward_std	None
kl	None
train_runtime (s)	99.1815

LoRA: r=64, alpha=64, QLoRA 4-bit training on an AMD Instinct MI325X (ROCm), default PEFT/Unsloth initialization, no warmup, learning rate 1e-4 (PLAN §7).