Searchformer offers the official code base for the paper “Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping.” This repository includes code for accessing datasets and trained models, reproducing paper figures, and running experiments. All code revolves around storing and transforming datasets in a MongoDB instance. The “notebook” folder contains Jupyter notebooks demonstrating how to access token datasets and prompt a trained Searchformer model. The “searchformer” module encompasses code for the experiments presented. Setup involves creating a Python 3.10 virtual environment with “requirements.txt” and setting up a MongoDB instance. Notably, the documentation provides comprehensive information on training models, evaluating them, and generating training datasets.
https://github.com/facebookresearch/searchformer