We’ve released Model Evolution, an open-source framework for recording machine-learning experiments. It keeps hypotheses and conclusions connected to the code, data, runs, and weights used in each experiment, including modules reused from earlier models. It is designed for researchers and engineers working with their existing training tools.
We built Model Evolution while developing models at Machine Learning Labs. Keeping track of code and checkpoints was straightforward; recording why we made each change and what we learned from it was harder. Reusing modules from earlier models also meant tracking where their weights came from, how they had been evaluated, and which combinations were compatible.
An experiment starts with a Markdown document describing the hypothesis, method, expected evidence, and what would falsify the claim. It keeps the same ID through planning, execution, and conclusion, with separate records for each run. Those records capture the code revision, configuration, dataset, inherited modules, results, and artifacts. The conclusion sits alongside the original plan, preserving how the evidence supported or rejected the idea.
Research notes and references live in Git. Datasets, weights, and full reports are published to an immutable artifact store, with checksums linking them to the records. Reusable modules have their own lineage and compatibility information, so an experiment can inherit weights from several earlier models while retaining a record of each dependency.
Agents make maintaining these records more practical. They can help write experiment plans, inspect recorded evidence, and document conclusions within a defined, validated structure. The repository includes an agent workflow for planning and concluding experiments; execution remains an explicit action by the researcher. A small adapter connects Model Evolution to the project’s existing code for dataset generation and training, without requiring a particular machine-learning framework.
Model Evolution is available under the Apache 2.0 license. The repository includes installation instructions and an example experiment.