Pretrained models for "Training Language Models for Social Deduction with Multi-Agent Reinforcement Learning" (AAMAS 2025 Version)