This repository provides resources and insights based on the paper Exploring Expert Failures Improves LLM Agent Tuning. The method introduced in this paper, termed Exploring Expert Failures (EEF), leverages beneficial actions from expert trajectories initially deemed unsuccessful, significantly improving the fine-tuning of Large Language Model (LLM) agents.

The expert trajectories dataset used in this paper is publicly available on Hugging Face at lclan/webshop_expert_trajectories. This dataset comprises trajectories generated by both GPT-4 and GPT-3.5-turbo within the WebShop environment, capturing both successful and failed attempts, enabling research into advanced agent tuning methods that benefit from analyzing and learning from expert failures.