ML-BENCH: EVALUATING LARGE LANGUAGE MODELS AND AGENTS FOR MACHINE LEARNING TASKS ON REPOSITORY-LEVEL CODE

Despite Large Language Models (LLMs) achieving impressive results in code generation, significant challenges remain in automated ML development, particularly in utilizing existing ML repositories effectively. Also, recently, people have developed LLM agents that attempt to interact with repository code (e.g., resolving issues), prompting the need for end-to-end evaluations starting from environment setup to deploying the repository rather than merely generating code in already-configured environments. These two gaps have motivated our development of ML-Bench, a benchmark rooted in real-world ML applications that leverage existing code repositories. ML-Bench encompasses annotated 9,641 examples across 18 GitHub repositories, challenging LLMs to accommodate user-specified arguments and documentation intricacies effectively. To evaluate both LLMs and agents, two setups are employed: ML-Bench-L for assessing LLMs' text-to-code conversion within a predefined deployment environment, and ML-Bench-A for testing autonomous agents in an end-to-end task execution within a Linux sandbox environment. Our findings indicate that while GPT-4o leads with a Pass@5 rate surpassing 50%, there remains significant scope for improvement, highlighted by issues such as hallucinated outputs and difficulties with bash script generation. Notably, in the more demanding ML-Agent-Bench, GPT-4o achieves a 76.47% success rate, reflecting the efficacy of iterative action and feedback in complex task resolution. Our resources, including code, data, and models, are available at \url{https://anonymous.4open.science/r/ML-Bench}.