Fine-tuned Llama-3.2-1B-Instruct via GRPO on the social-agent-negotiation-v1 OpenEnv environment.

Training: 3 epochs × 8 episodes × 3 tasks (single-round-consensus, adversarial-information, opioid-overdose)
Method: Group Relative Policy Optimization (GRPO) via HuggingFace TRL
Base model: unsloth/Llama-3.2-1B-Instruct (4-bit LoRA, r=16, alpha=32)

See the training notebook: training/grpo_training.ipynb in the GitHub repo.