trainable-agents

One Minute Overview#

Character-LLM is a trainable agent specifically designed for role-playing that can think and respond like historical figures such as Beethoven, Cleopatra, and Socrates. Unlike simple prompted agents, Character-LLMs are specially trained with detailed character-specific knowledge and representative personality traits, eliminating the need for constant prompting.

Core Value: Generates diverse training data through Experience Reconstruction technology, enabling AI to authentically simulate the behavior, language, and knowledge systems of historical figures.

Quick Start#

Installation Difficulty: High - Requires GPU infrastructure and specialized AI knowledge

# Clone repository
git clone https://github.com/choosewhatulike/trainable-agents.git
cd trainable-agents

# Install dependencies
pip install -r requirements.txt

# Download model weights
# Requires base model (llama-7b) first, then apply weight differences

Is this suitable for me?

• ✅ Role-playing game development: Scenarios requiring interaction with historical figures
• ✅ Educational applications: Letting students learn by conversing with historical figures
• ✅ Creative writing assistance: Providing dialogue and interactions in specific character styles
• ❌ Simple chatbots: No need for specific role-playing functionality
• ❌ Commercial applications: Models and data are for academic research only

Core Capabilities#

1. Authentic Role-Playing - Accurate Historical Figure Simulation#

Character-LLMs can think and respond like specific historical figures, using tone, manner, and vocabulary appropriate to each character, with complete knowledge of that character. Real Value: Creates genuinely lifelike historical character interaction experiences, enhancing authenticity in educational and entertainment applications

2. Character-Specific Knowledge Base - Detailed Character Backgrounds#

Each character model contains extensive training data specific to that character, averaging 1.6K scenes, 754K words, and 13.2 dialogue turns. Real Value: Ensures character responses align with historical facts and personality traits, preventing responses inconsistent with the character's identity

3. Experience Reconstruction Technology - High-Quality Data Generation#

Through a unique Experience Reconstruction process, uses GPT-3.5-turbo to generate detailed and diverse character experience data, including scene extraction, interaction generation, and hallucination protection scenarios. Real Value: Reduces manual data collection costs while improving character model quality and consistency

4. Multi-Character Support - Nine Historical Figures#

Offers pre-trained models for nine historical figures including Beethoven, Cleopatra, Voldemort, Spartacus, Hermione, Newton, Caesar, Socrates, and Martin Luther King Jr. Real Value: No need to train from scratch, directly use high-quality historical character models

5. Distributed Training Support - Efficient Model Training#

Supports multi-GPU distributed training, completing model training in 30-45 minutes on 8 A100 GPUs. Real Value: Accelerates training of large-scale character models, improving research efficiency

Technology Stack & Integration#

Development Language: Python Main Dependencies: FastChat, Transformers, Hugging Face Hub, PyTorch Integration Method: API / SDK / Library

Ecosystem & Extension#

• Plugins/Extensions: Based on FastChat framework, supports extended functionality
• Integration Capabilities: Can integrate with existing LLM ecosystems, supports model weight differential application

Maintenance Status#

• Development Activity: Active project with regular updates
• Recent Updates: Recent maintenance activities supporting model training and inference
• Community Response: Related discussions in academic research community, suitable for research and experimentation

Commercial & License#

License: Code Apache-2.0, Data CC BY-NC 4.0

• ✅ Commercial: Code allows commercial use, models and data for academic research only
• ✅ Modification: Allows code and model modification
• ⚠️ Restrictions: Models and data cannot be used for commercial purposes

Documentation & Learning Resources#

• Documentation Quality: Comprehensive
• Official Documentation: https://github.com/choosewhatulike/trainable-agents
• Example Code: Provides single-turn and multi-turn conversation examples, character creation tutorials