--- title: werewolf_lite canonical_url: "https://www.modelscope.ai/models/SKYLENAGE/werewolf_lite" md_url: "https://www.modelscope.ai/models/SKYLENAGE/werewolf_lite.md" repository: SKYLENAGE/werewolf_lite last_updated: 2026-06-24 library_name: - pytorch frameworks: - pytorch downloads: 76 stars: 0 --- # werewolf_lite > werewolf_lite - An open-source model by SKYLENAGE on ModelScope. Werewolf Lite — 9-Player AI Werewolf Engine - **Repository**: SKYLENAGE/werewolf_lite - **Downloads**: 76 - **Stars**: 0 - **Last updated**: 2026-06-24 Source: https://www.modelscope.ai/models/SKYLENAGE/werewolf_lite --- # Werewolf Lite — 9-Player AI Werewolf Engine **English** | [中文](README_zh-CN.md) The **AI Werewolf Arena** project of [Skylenage](https://skylenage.net) [Game Arena](https://skylenage.net/sla/arena/gameArena) — a lightweight environment for running large language models as players in a standard 9-player Werewolf (Mafia) game, covering reasoning, discussion, voting, and role-based interactions. ## Features - **Standard 9-Player Setup**: 3 Werewolves / 1 Seer / 1 Witch / 1 Hunter / 3 Villagers - **Multi-Model Support**: Unified access to OpenAI / Anthropic / DashScope and other LLMs via [LiteLLM](https://github.com/BerriAI/litellm) - **Core Game Flow**: Bidding-based discussion, simultaneous majority voting, sheriff election, witch save/poison, hunter shoot, last words - **TTS Voice Playback**: Independent voice timbre for each player, powered by DashScope CosyVoice - **LLM Narrative Summary**: Per-cycle recap and post-game review commentary ## Quick Start > **Requirements**: Python >= 3.10 ### 1. Install Dependencies ```bash pip install -r requirements.txt ``` ### 2. Configure API Keys Edit `configs/nine_player_chinese.yaml` and fill in the LLM API info for each player: ```yaml game_config: agents: - role: "Werewolf" id: "Player1" agent_id: "chinese_llm" display_name: "GPT-4o" chat_mode: "text" llms: - model_name: "openai/gpt-4o" api_base: "https://api.openai.com/v1" api_key: "sk-your-api-key-here" ``` > See `configs/nine_player_chinese.yaml` for the full 9-player template. ### 3. Run the Game ```bash python run.py --config configs/nine_player_chinese.yaml ``` Main CLI arguments: | Argument | Description | Default | |----------|-------------|---------| | `--config` / `-c` | Game config file | `configs/nine_player_chinese.yaml` | | `--output` / `-o` | Output directory | `./outputs` | | `--debug` / `-d` | Enable debug mode | off | | `--random` / `-r` | Use random agents for quick testing | off | | `--shuffle` / `-s` | Randomly shuffle role assignment | off | > **Random Seed**: Set via `game_config.seed` in the config file. Leave empty for random each run. ## Optional Features ### TTS Voice Synthesis Requires a [DashScope](https://dashscope.aliyuncs.com) API key. Edit `tts_config.yaml`, then run: ```bash python werewolf_tts.py -g outputs/werewolf_game_xxx.json -o ./audio_output ``` ### Narrative Summary Generation Requires an OpenAI-compatible API. Edit `summary_config.yaml`, then run: ```bash # Single game summary python werewolf_summary.py outputs/werewolf_game_xxx.json # Batch processing python werewolf_summary.py --batch --skip-existing ``` ## Project Structure ``` werewolf_lite/ ├── game/ # Core game engine │ ├── engine.py # Main controller (Moderator FSM) │ ├── states.py # Game state management │ ├── roles.py # Role definitions │ ├── actions.py # Player action definitions │ ├── records.py # Event & data records │ ├── consts.py # Constants & enums │ ├── base.py # Game base classes │ ├── night_elimination_manager.py # Night resolution manager │ └── protocols/ # Discussion / voting / bidding protocols │ ├── bid.py # Bidding protocol (incl. UrgencyBidding) │ ├── chat.py # Discussion protocol │ ├── vote.py # Voting protocol │ ├── ordering.py # Speaking order │ └── factory.py # Protocol factory ├── harness/ # AI agent integration layer │ ├── base.py # Agent base class & LLM invocation │ ├── chinese_agent.py # Chinese LLM agent │ └── litellm_models.yaml # Model pricing config ├── configs/ # Game configs │ └── nine_player_chinese.yaml ├── run.py # Game entry point ├── player_beliefs_analysis.py # ToM six-dimension scoring tool ├── werewolf_tts.py # TTS voice synthesis tool ├── werewolf_summary.py # LLM narrative summary tool ├── tts_config.yaml # TTS config ├── summary_config.yaml # Summary config └── requirements.txt ``` ## ToM Scoring Analysis A six-dimension Theory-of-Mind scoring scheme based on `player_beliefs`, offering a quantitative view of LLM reasoning behavior in social-deduction games: | Dimension | Description | |-----------|-------------| | First-Order Reasoning | Accuracy of judging others' identities | | Second-Order Recursion | Accuracy of predicting "how others perceive me" | | Belief Stability | Consistency of judgments across rounds | | Dynamic Update | Efficiency of correcting errors after key information | | Strategic Alignment | Villager truth-convergence / Werewolf deception-success | | Game Depth | Temporal stability of second-order judgments | ```bash # Single-game analysis python player_beliefs_analysis.py -i outputs/werewolf_game_xxx.json # Batch analysis python player_beliefs_analysis.py -i outputs --batch ``` ## Configuration Reference ### Game Config (`configs/nine_player_chinese.yaml`) | Field | Description | |-------|-------------| | `game_config.randomize_roles` | Whether to randomly assign roles | | `game_config.discussion_protocol` | Discussion protocol (bidding/sequential) | | `game_config.day_voting_protocol` | Daytime voting protocol | | `agents[].llms` | Per-player LLM config (model, API base, key) | ### TTS Config (`tts_config.yaml`) | Field | Description | |-------|-------------| | `default.api_key` | DashScope API key | | `default.model` | Voice model (default: `cosyvoice-v3-plus`) | | `players..voice` | Independent voice timbre per player | ## Roadmap - **Diversified Game Setups**: Support varying player counts, role compositions, and rule variants (e.g. Guard, Idiot, Wolf King) to observe LLM behavior under different information-asymmetry structures and game complexities - **Human-AI Mixed Play**: Allow human players to participate alongside AI agents, exploring interaction patterns between humans and LLMs in social-reasoning tasks - **Tournament & Ranking System**: Introduce Elo ratings and multi-game statistics as a reference for model-to-model comparison - **Dynamic Strategy Evolution**: Explore game-feedback-driven iteration of agent strategies to further study ToM reasoning ## License MIT License — see [LICENSE](LICENSE)