Qwen3.8-27B GRPO metasploit defender
Stage-5 GRPO (Unsloth + TRL) LoRA adapter for the Qwen3.8-27B cyber-range defender policy, focused on the metasploit tool family.
- Run:
metasploitround 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) | 81.2435 |
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).
- Downloads last month
- 17
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3.8-27B-unsloth-bnb-4bit") model = PeftModel.from_pretrained(base_model, "brosnanyuen/Qwen3.8-27B-GRPO-metasploit-defender")