Improving Tool Acquisition in Large Language Models

Supporting material for the thesis "Thinking Forwards, Backwards, and in Code: Improving Tool Acquisition in Large Language Models"

Authors: Huey Sun

Repository Structure

• Results: The ToolSandbox evaluations results can be found here. Each result has a summary at the end, which was used to create the tables in the thesis.

• Data: The data used to finetune the models can be found here. As Mistral's lack of chat formatting means that incremental masking is impossible (the output is not perfectly autoregressive at each conversation turn), each assistant message has been individually preformatted.

• Scripts: The scripts used to convert json conversations to Mistral's expected formatting can be found here.

Other Code

You can find my fork of the torchtune library here, where I add support for tool chat and Mistral formatting in training.