LMGame Training Framework

A comprehensive framework for multi-turn reinforcement learning training of language model agents in gaming environments.

Quick Start

Prerequisites

• Python 3.10
• CUDA-compatible GPU (A100, L40, or similar)
• Conda package manager

Installation

1. Create conda environment:

   conda create --name lmgame_train python=3.10
   conda activate lmgame_train

2. Set up authentication (optional but recommended):

   export WANDB_API_KEY=your_wandb_api_key
   export WANDB_ENTITY=your_wandb_entity
   export HF_TOKEN=your_huggingface_token

3. Run setup script:

   ./scripts/setup.sh

Training Examples

Sokoban Agent Training

source train_sokoban.sh

Hardware Configuration

The framework is pre-configured for different GPU setups:

GPU Type	Agent Groups	Group Size	Total Agents	Default Model
A100 (default)	8	16	128	Qwen/Qwen2.5-0.5B-Instruct
L40	4	2	8	Qwen/Qwen2.5-0.5B-Instruct

Note: The A100 configuration is the default setting in configs/base.yaml. For other GPUs, adjust agent_group_num and agent_group_size in the config file.

Documentation

• System Design Overview - Architecture and design principles
• Development Guide - Contributing and development workflow

License

This project is licensed under the MIT License - see the LICENSE file for details.