adaraai/adara-evals

African AI evaluation harness and benchmark architecture # adara-evals African AI evaluation and benchmarking architecture (no fabricated scores). ## Status **Active (harness only)** Classification: `EVALUATION` · Visibility: `public` · License: `Apache-2.0` ## Overview Defines the ADARA African AI Benchmark layout: language, speech, codeswitch, culture, reasoning, safety. Runners record dataset version, model, methodology, metrics, config, timestamp, and limitations. **No benchmark results are checked in.** ## Why this exists Evaluation must be independent of training and product code. ## Architecture benchmarks/*/ for task defs; runners/ for execution; reports/ for generated output (gitignored when they contain scores you have not published). ``` AFRICAN DATA ↓ LANGUAGE / SPEECH / CONTEXT ↓ MODELS ↓ EVALUATION + SAFETY ↓ API → SDK → PRODUCTS ``` See repository map in `adara-platform`. ## Installation Python 3.12. `pytest -q` ## Usage See `schemas/run.json`. Do not invent leaderboard numbers. ## Development Use Conventional Commits and pull requests against `main`. See CONTRIBUTING.md. ## Testing Run the repository's documented test command. Do not run expensive training on every PR. ## Roadmap Documented in this README's status and in `adara-platform/ADARA-ARCHITECTURE.md`. Do not treat planned work as shipped. ## Contributing See CONTRIBUTING.md and the organization profile. ## Security See SECURITY.md. Never commit secrets, model weights, or private datasets. ## License Apache-2.0