AIDE ML

AIDE ML is the open-source reference implementation of the AIDE algorithm, developed by Weco AI. It is a Python package in which an LLM agent drafts, debugs and benchmarks machine learning code until a user-defined metric is optimised. The latest listed release is v0.2.2.

What It Is

AIDE ML is a machine learning engineering agent. A user points it at a dataset and describes the goal and evaluation metric in plain English. The agent then runs a tree search over code, where each script is a node and LLM-generated patches create child nodes. Metric feedback guides and prunes the search. The repository positions it for researchers who want to replicate the paper, test new search heuristics or evaluators, and for ML practitioners prototyping pipelines.

Workflow and Outputs

Runs are started with the aide CLI, from Python via aide.Experiment, or through a Streamlit web UI. The UI shows live logs, the solution tree and the best code. After a run, the best code is saved to best_solution.py and an HTML tree plot lets you inspect the code attached to each node. Settings such as the number of steps and drafts per step can be set through command-line flags.

Model Access and Deployment

The agent works with OpenAI, Anthropic, Gemini and OpenRouter, and with local LLMs that speak the OpenAI API, such as Ollama. A fully local setup is documented, though the README notes some performance drop with fully local models. A Dockerfile is provided for containerised runs. The code is MIT licensed.

Relation to Weco

The README presents the Weco platform as the product that generalises AIDE's capabilities to broader code optimisation, adding experiment tracking and more user control. The README also lists research projects from several institutions that build on or with AIDE.