🕵️ Among Agents — Qwen2.5-1.5B Fine-Tuned

A language model fine-tuned to play social deduction games — navigating, deceiving, and reasoning about other agents' beliefs in real time.

🧠 What Is This Model?

• 🗺️ Navigate a 14-room spaceship map as a crewmate or impostor
• 🧩 Complete tasks while tracking other agents' movements
• 🗳️ Reason and vote during emergency meetings
• 🎭 Deceive or deduce — depending on assigned role

📊 Training Results

Training Loss Curve

Loss dropped 97.5% over 800 steps — rapid convergence in the first 50 steps, stable improvement thereafter

Episode Reward Distribution

Mean reward of 19.13 across 100 episodes — bimodal distribution shows the model learned distinct crewmate and impostor strategies

⚡ Before vs. After Fine-Tuning

Base Model (Qwen2.5-1.5B)

Observation: You are Yellow. You are in Cafeteria.
Action: I think I should go somewhere and do something useful maybe look around

Fine-Tuned Model

🏋️ Training Configuration

🌍 Environment

• 7 simultaneous agents (6 crewmates + 1 impostor)
• 14-room Skeld map with pathfinding and collision
• Partial observability — each agent sees only their current room
• Dense composable rewards (+0.30 task completion, +1.00 correct vote, −2.00 killed)
• Two game phases: Exploration and Emergency Meeting

🔗 Resources

📜 Citation