GenerativeAgentsCN

Simplified Chinese | English

Generative Agents (Generative Agents) in-depth Chinese version

Stanford AI Town was open sourced by Stanford University and Google in August 2023. It is a virtual world composed of 25 agents that simulates real human life.

25 agents are completely driven by ChatGPT and can independently organize parties, attend meetings, and plan various activities for Valentine's Day. They can display life patterns and behavioral habits similar to humans.

The original code of Generative Agents has a low degree of engineering, making it difficult to maintain or expand its functions continuously. Moreover, after more than a year, the capabilities of Chinese LLM are already qualified for such tasks. Therefore, we reconstructed and in-depth localized the original project, aiming to provide Chinese users with a basic version that is easy to maintain for subsequent experiments or trying more gameplay.

The wounderland project is a reconstructed version of the original Generative Agents project. It has a good structure and the code quality is much better than the original version. Therefore, this project is developed based on wounderland.

Main work:

• Rewrite all prompts and switch the agent's "native language" to Chinese to facilitate docking with Chinese models such as Qwen-2.5 or GLM-4;
• Based on the characteristics of Chinese and the capabilities of the Qwen2.5-7b/14b model, optimize the Chinese prompts and the start and end logic of dialogue between agents;
• All prompts are templated to facilitate later maintenance;
• Fixed minor problems in the original version (for example, the agent in the original version of wounderland no longer wakes up after falling asleep);
• Add support for the local Ollama API, and also connect LlamaIndex embedding to Ollama to achieve complete local deployment and reduce experimental costs. For Ollama installation and configuration, please refer to ollama.md ;
• Added features such as "breakpoint recovery";
• The playback interface is simplified based on the original Generative Agents front-end code, and the timeline of agent activities and dialogue content are saved to Markdown documents.

Playback screen:

Note: The map and character names are also synchronized into Chinese to prevent LLM from switching to English context when encountering mixed Chinese and English contexts.

1. Preparation

1.1 Get the code:

 git clone https://github.com/x-glacier/GenerativeAgentsCN.git
cd GenerativeAgentsCN

1.2 Configure large language model (LLM)

Modify the configuration file generative_agents/data/config.json :

1. Ollama is used by default to load local quantization models and provides OpenAI compatible APIs. You need to first pull the quantitative model (refer to ollama.md), and ensure that base_url and model are consistent with the configuration in Ollama.
2. If you want to call other manufacturers' APIs, you need to fill in the corresponding key in api_keys and modify base_url and model according to the API documentation.

1.3 Install python dependencies

It is recommended to use anaconda3 to create and activate the virtual environment first:

 conda create -n generative_agents_cn python=3.11
conda activate generative_agents_cn

Install dependencies:

 pip install -r requirements.txt

2. Run a virtual town

 cd generative_agents
python start.py --name sim-test --start "20240213-09:30" --step 10 --stride 10

Parameter description:

• name - Each time you start the virtual town, you need to set a unique name for later playback.
• start - the starting time of the virtual town.
• resume - After the run ends or is unexpectedly interrupted, continue running the virtual town from the last "breakpoint".
• step - the number of iteration steps after which to stop running.
• stride - the time (minutes) corresponding to each iteration in the virtual town. If --stride 10 is set, the time changes of the virtual town during the iteration process will be 9:00, 9:10, 9:20...

3. Playback

3.1 Generate playback data

 python compress.py --name <simulation-name>

After the operation is completed, the playback data file movement.json will be generated in the results/compressed/<simulation-name> directory. At the same time, simulation.md will be generated to present the status and dialogue content of each agent in a timeline.

3.2 Start playback service

 python replay.py

Open the replay page through the browser (address: http://127.0.0.1:5000/?name=<simulation-name> ), and you can see the activities of the residents in the virtual town in various time periods.

You can move the screen with the arrow keys

Parameter description

• name - the name set when starting the virtual town.
• step - the starting step number of playback, 0 means starting playback from the first frame, the default value is 0.
• speed - playback speed (0-5), 0 is the slowest, 5 is the fastest, and the default value is 2.
• zoom - the screen zoom ratio, the default value is 0.8.

Replay data named example (generated by qwen2.5:32b-instruct-q4_K_M) is built into the release version. If you want to start playback from the beginning at a faster speed and the screen zoom ratio is 0.6, the corresponding url is: http://127.0.0.1:5000/?name=example&step=0&speed=2&zoom=0.6

You can also open simulation.md directly to view all character activities and dialogue information in example .

3.3. Playback screenshots

The dialogue content in the screen is generated by qwen2.5:14b-instruct-q4_K_M

town panorama

garden

cafes

classroom

4. References

4.1 Paper

Generative Agents: Interactive Simulacra of Human Behavior

4.2 Code

Generative Agents

wounderland