3l3ktr4/donorsim-qwen3-8b-abstract-step20
The 3l3ktr4/donorsim-qwen3-8b-abstract-step20 model is an 8 billion parameter Qwen3-based language model, fine-tuned using GRPO for simulating and understanding iterated Donor's Game scenarios. It processes naturalistic descriptions of social interactions and predicts choices based on normalized payoff and reciprocity, with partners rotating and memory maintained. This model specializes in abstract social reasoning within game theory contexts, distinguishing it from general-purpose LLMs.
Loading preview...
Model Overview
3l3ktr4/donorsim-qwen3-8b-abstract-step20 is an 8 billion parameter Qwen3 model that has undergone specialized fine-tuning using the GRPO method (verl 0.7.1, LoRA r16/alpha32 merged into bf16 weights). Its primary focus is on simulating and analyzing the abstract / everyday-scene stage of the iterated Donor's Game.
Key Capabilities
- Social Interaction Simulation: Processes short, naturalistic descriptions of situations involving named individuals in a small group.
- Choice Prediction: Outputs
CHOICE: 1orCHOICE: 2based on the scenario, with randomized option order. - Reward System Integration: Rewards are calculated based on normalized payoff (Term 1) and reciprocity (Term 2), without explicit group or CFE terms.
- Partner Rotation and Memory: Simulates partner rotation within a roster of
n_players - 1members, maintaining per-partner memory. - Contextual Understanding: Interprets re-encounter probability (
w) and gossip probability (q) phrased in words each turn.
Training Details
This model received 20 abstract steps of fine-tuning, building upon donorsim-qwen3-8b-modeAB-step75, which itself had 75 steps of structured group game training. The full weights are merged, allowing direct loading with transformers or vLLM without requiring an adapter.