BEATS: OPTIMIZING LLM MATHEMATICAL CAPABILITIES WITH BACKVERIFY AND ADAPTIVE DISAMBIGUATE BASED EFFICIENT TREE SEARCH

Answer Generation

python /generate/treeSearch.py

Before processing, set your own base LLM, load_data_path, save_path.

In this version of code, we only provide prompts for LLaMA and we will update prompts for Qwen as soon as the paper is accepated.

The prompt is located in /generate/action_prompt_3_llama.py

Eval

The eval framework are based on excellent work MAmmoTH and we made some improvements.

To use the eval tools:

python /eval/math_eval/run_open_mcts.py

For majority voting, please use function eval_base().

For back verify, please use function eval_backVerify().