Future of Society Public Goods Game Dataset

This dataset contains the analysis-ready decision log from the Public Goods Game (PGG) case study reported in "Future of Society: A Platform for Controlled Behavioral Experiments with Mixed-LLM Agent Populations" (Anonymous ACL submission). The experiment ran a repeated linear voluntary contribution mechanism with 200 LLM-powered agents per profile-generation context across five profile-generation contexts, three SimTree multiplier branches (1.3×, 2.0×, 5.0×), and ten rounds per branch, yielding 30,000 planned decisions. This file contains the 29,590 decisions retained after quality filtering. Each row is one agent decision and includes:- Agent identity and persona (role, profile text, archetype)- Demographic fields: Age (18–30, 31–50, 51+) and Income (Low, Middle, High)- Continuous traits: Trust and Risk Tolerance (Gaussian draws, mean 50, std 15)- Profile-generation context (profile_llm): the LLM used to write the agent's profile — one of Gemma, Granite, Ministral, Phi4, Qwen- Decision-making LLM (llm_model): the model assigned to make this agent's decisions — gemma3:4b-it-qat, granite4:3b, ministral-3:3b, phi4-mini:latest, qwen3:4b-instruct-2507-q4_K_M- Multiplier branch (1.3, 2.0, 5.0) and round (1–10)- Action taken (allocate / keep), contribution amount, and follow-up amount- Data quality classification: clean (29,528), amount_over20 (58), action_variant (4)- Derived variables: contribution (0 if keep), chose_allocate, capped flag The platform code is available at:https://github.com/ZJU-Computational-Social-Science-Lab/Future-of-Society Decoding parameters: Ollama 0.8 temperature, topp 0.9, topk 40. No fixed random seed was used; five independent stochastic runs were conducted to demonstrate finding stability across stochastic variation. Hardware: Intel Core i9-14900HX, 32GB RAM, NVIDIA RTX 5070 Laptop GPU, Windows 11, Ollama 0.23.2. Total runtime: 23.3 hours. API calls: 0.