ML Configuration System

A 1-of-N research-track option. Design and partially implement an ML configuration system — schemas, inheritance/overrides, validation, and reproducibility. Specifics of the task have not been publicly documented; treat this card as the structural shell. Requirements The round is one of the four to six choices on the research-track menu (alongside Coding & Design, ML Programming, Prompting & Engineering with LLMs, RL Fundamentals, and the ML Take-Home). Reported expectations: Design a config representation that supports composition (mixins/inheritance), overrides at the CLI/API layer, and validation against a schema. Reason about reproducibility: how do you guarantee that a training run can be re-launched from a frozen config, weeks later, on a different node? Trade off ergonomics (terse configs) vs. safety (strict schemas). Some candidates report a coding component ("flesh out the loader"); others report a discussion-only round. Notes The full prompt has not been publicly documented; candidates who explicitly asked for more detail did not get it. Expect overlap with how Hydra, OmegaConf, or pydantic-settings work; mention them by name. Preparation Read Hydra's documentation and reproduce its config-composition story in plain Python: base configs, group overrides, command-line interpolation. Be able to discuss pydantic for typed validation and the tradeoff vs. dynamic dataclasses. Think through what "reproducibility" actually requires (hashing the resolved config, freezing the env, version-pinning data).