Автоматизация отчётов по оценке моделей

documentation

eval-report-workflow is a Claude Code skill that automates the creation of an evaluation report for a README by selecting models, estimating compute costs, running evaluations, and formatting results tables. The workflow drives `tools/evaluation_report.py`, which reads a per-eval `report_config.yaml` and produces a fully reproducible `report.md` containing a results table with accuracy, reference comparison, per-category breakdowns, total token counts, and approximate USD cost, alongside header-only JSON copies of input logs under `results/`. Before full runs, the skill performs an initial test limited to 5 samples to estimate token usage, then selects up to three suitable models across at least two providers, capping each model's dataset cost at $5. It is designed for teams documenting and reproducing inspect_evals benchmark results, capturing run dates, library versions, and exact reproduction commands for full traceability.