worker_experiment

Your Task

Before doing anything, create a log file ${exp_result_dir}/experiment_log.md, refer to the blueprint to get ${exp_result_dir}.

1. Run the experiment according to the experiment blueprint.
   • when execute training commandline, record your command into ${exp_result_dir}/experiment_log.md.
2. Wait for the experiment to finish, error or time out.
3. If the experiment fails, attempt to fix it and document the debugging process (${exp_result_dir}/experiment_log.md). If it cannot be fixed, skip to step 5.
4. Place comprehensive experiment results in the designated location (${exp_result_dir}/experiment_log.md).
5. Create a finish.flag file in [exp_result_dir] to mark the task as complete.
6. Done.

Experiment Blueprint:

Experiment blueprints are designed to execute experiments that validate hypotheses or gather necessary data. An experiment blueprint is a markdown file (blueprint.md). It contains 7 sections — you should execute the task based on this information:

1. [exp_purpose] Experiment purpose (text): Briefly describe the main purpose of this experiment and the key differences from other blueprints (e.g., which hyperparameter or environment variable differs).
2. [exp_codebase_dir] Main experiment code path (absolute path): The absolute path containing all code needed to run the experiment. Relatively small in size. Does not include the Python virtual environment. Remember to cd to this path before starting the experiment.
3. [exp_venv_exe] Python virtual environment path (absolute path to python executable): Path to the Python executable. E.g.: /mnt/data_cpfs/agentjet/project/.venv/bin/python
4. [exp_yaml_path] Experiment config file path (absolute path) (may not exist in swarm training mode): Path to the experiment configuration YAML file. This file must be located within the main experiment code path. E.g.: /mnt/data_cpfs/agentjet/project/tests/bench/benchmark_math/benchmark_math.yaml
5. [exp_launch_command] Training execution command (string): E.g. python -m ajet.launcher --conf tests/bench/benchmark_math/benchmark_math.yaml --skip-check-avail-gpu --with-ray
6. [exp_result_dir] Result data storage path (absolute path): Path for output data storage
7. [exp_max_time] Maximum runtime is ${MaxTime}; each experiment is forcefully terminated after ${MaxTime}