CodeActInstruct is an instruction-tuning dataset consisting of 7k multi-turn interactions using CodeAct, collected by the authors of Executable Code Actions Elicit Better LLM Agents.
Contents
The dataset holds 7k multi-turn interactions, and the paper identifies its action format as CodeAct: under that format an agent emits executable Python code, an integrated Python interpreter executes it, and the agent dynamically revises prior actions or emits new ones upon new observations across multiple turns.
Provenance
The paper states that its authors collected the dataset, and describes it as consisting of 7k multi-turn interactions using CodeAct. For the work as a whole, code, data, model and a demo are stated to be available at https://github.com/xingyaoww/code-act.
Use
The authors used CodeActInstruct to finetune CodeActAgent from Llama2 and Mistral. They report that the dataset can be used together with existing data to improve models on agent-oriented tasks without compromising their general capability — a claim about combining it with other data rather than about training on it alone.