This directory contains various scripts for metaplan generation, sampling, and evaluation.

Script for generating metaplans. Supports asynchronous batch processing for efficiency.

Constructs preference pairs based on metaplan execution success rates. This script analyzes the execution results of different metaplans and selects the best and worst performing metaplans as preference pairs.

Splits generated metaplan samples into multiple files for parallel evaluation.

Main script for metaplan quality evaluation. It coordinates the entire workflow:

1. Sampling multiple metaplans
2. Launching task completion agents
3. Evaluating metaplan quality
4. Constructing preference pairs

Contains prompt templates for ALFWorld and SciWorld environments. These templates guide large language models to generate structured metaplans.

A typical workflow is:

1. Use gen_metaplan.py to generate multiple metaplans
2. Use split_metaplans_for_sample.py to split metaplans into multiple files
3. Use construct_metaplan_pairs.py to construct preference pairs

For detailed parameters and options, please refer to the help documentation of each script.