6. The full source code and the models used

Purpose of this post

This post publishes the source code used for the Tetris AI. The code is loaded only when you open a file.

Models used

• stable-baselines3 DQN: the low-level 7-action models (300,000 and 370,000 steps) and the macro 40-action model (300,000 steps). The model files are not included.
• Qwen2.5-3B-Instruct GGUF q4_k_m (llama-cpp-python): text board judge, picking a candidate number
• Qwen2.5-VL-3B-Instruct (transformers): image board judge, picking spots directly

Folder layout

• tetris: C++ engine (tetris_engine), C API (tetris_c_api), Qt6 screen and TCP server (main_window, tetris_widget), CMake
• local_tetris_env.py: gymnasium environment wrapping the engine with ctypes
• simple_tetris_env.py: a pure Python environment used when the engine is not built. Its observation and action formats differ, so models cannot be swapped between them.
• macro_planning_env.py: the 40-action environment that chooses once per piece
• train_tetris_rl.py: DQN training without a screen
• train_qt_live.py: shows the current model on the Qt screen while training. It also has a mode that trains on games played by a person.
• qt_rl_client.py: attaches a trained model to the Qt screen
• llm_teacher.py, vlm_teacher.py, llm_reward_wrapper.py: LLM and VLM reward judges
• vlm_macro_client.py: client where the VLM picks spots directly
• llm_macro_client.py: candidate computation and LLM number picking
• moe_macro_policy.py: candidate selection mixing four rules
• config/rl_config.py: training settings

Build outputs, virtual environments, model weights and trained model files are excluded.

Full source