What problem does it solve? Building a benchmark for a new data domain (spectroscopy, genomics, time series) normally means reimplementing model zoos, cross-validation splitting, experiment runners, and leaderboard statistics from scratch. This Skill guides you through reusing TabArena's model registry, task runner, and bencheval's leaderboard math from your own external repository instead of forking or reimplementing that layer. ## Core Features & Use Cases - Task and Split Construction: Convert domain datasets into UserTask objects with repeated k-fold, group-aware, or temporal outer splits, plus inner bagging protocols via ValidationProtocol. - Model Registry Extension: Register domain-specific models with register_model_info alongside TabArena's zoo, or wrap self-tuning pipelines as ExternalSystemModel baselines. - Leaderboard and Publishing: Compute Elo, win-rates, and improvability with bencheval.BenchmarkEvaluator, handle imputation and Elo anchoring, and host results with MethodMetadata artifact tiers. - Use Case: A research lab building a Raman spectroscopy benchmark (RamanBench) uses this Skill to depend on tabarena and bencheval, register a spectral CNN model, define group-aware splits per specimen, and publish a leaderboard comparable to TabArena's. ## Quick Start Ask the assistant to adapt TabArena as the tasks, models, and leaderboard layer for your new domain benchmark, naming your benchmark and its data domain.
