MARS-SQL training

This directory contains the reinforcement-learning implementation and the mars-train.sh entry point used for MARS-SQL training. The vendored SkyRL-OpenHands/ and verl/ trees provide the agent environment and trainer runtime; project-specific launch settings live in mars-train.sh.

Setup

Follow Install.md to create the Python 3.12, CUDA 12.4, uv, and Ray environment. The launch script can be called from any working directory and uses the repository's top-level data/ directory by default.

Required configuration

Variable	Purpose
DB_PATH	Root containing BIRD database directories such as <db_id>/<db_id>.sqlite
CKPT_PATH	Parent directory in which checkpoints are written
WANDB_API_KEY	W&B credential used by the configured training logger

Optional variables include DATA_DIR, CUDA_VISIBLE_DEVICES, PROJECT_NAME, and EXPERIMENT_NAME. Set RESET_EXPERIMENT=1 only when the existing checkpoint directory for the selected project and experiment should be removed. Set RESET_RAY=1 to stop an existing local Ray runtime before launch.

Launch

export DB_PATH=/absolute/path/to/bird/databases
export CKPT_PATH=/absolute/path/to/checkpoints
export WANDB_API_KEY=your_wandb_api_key
bash Mars-train/mars-train.sh

The script validates its database directory and the required training Parquet files before starting Ray/verl, so path errors fail early with a useful message.