🎯 GameTheory Chat
AI-Powered Strategic Reasoning · Powered by Qwen2.5-7B + LoRA Fine-Tuning
Describe any real-world strategic scenario and the Formulator model will frame it as a game theory problem, solve it, and interpret the results.
Describe a strategic scenario...
Paste a formal game theory problem (payoff matrix, game description, etc.) and the Solver model will find equilibria and optimal strategies.
About GameTheory Chat
GameTheory Chat is the interactive demo for the GameTheory-Bench project — a 3-phase pipeline that fine-tunes Qwen2.5-7B-Instruct into a game theory specialist.
The Three-Phase Pipeline
| Phase | Model | Method | Result |
|---|---|---|---|
| Phase 1 — Solver | GameTheory-Solver | Supervised Fine-Tuning on 2,913 verified problems | 82% to 94% accuracy |
| Phase 2 — Reasoner | GameTheory-Reasoner | GRPO reinforcement learning (750 steps) | +6% reasoning quality |
| Phase 3 — Formulator | GameTheory-Formulator-Model | SFT on 1,215 real-world formulation problems | 100% valid formulations |
HuggingFace Assets
GameTheory-Bench
2,913 computationally verified game theory problems across 8 categories
View DatasetGameTheory-Formulator
1,215 real-world to game theory formulation problems (6 domains, 33 subtypes)
View DatasetGameTheory-Formulator-Model
Phase 3 SFT LoRA — translates real-world scenarios into formal models
View ModelBenchmark Results
| Category | Base Model | After Phase 1 (SFT) | After Phase 2 (GRPO) |
|---|---|---|---|
| 2x2 Normal Form | 78% | 95% | 96% |
| NxM Normal Form | 65% | 89% | 92% |
| Zero-Sum Games | 80% | 96% | 97% |
| Bayesian Games | 52% | 85% | 90% |
| Extensive Form | 58% | 88% | 92% |
| Mechanism Design | 45% | 82% | 88% |
| Cooperative Games | 60% | 90% | 93% |
| Evolutionary Games | 55% | 87% | 91% |
| Overall | 62% | 89% | 92% |
Technical Details
- Base model: Qwen2.5-7B-Instruct
- Training: QLoRA (r=32, alpha=64, 4-bit NF4 quantization)
- Hardware: Dual RTX 3090 (training), ZeroGPU A10G (inference)
- Inference: bfloat16, streaming generation
Built by Alogotron | Powered by Qwen2.5 + PEFT + Gradio