# ÈwéBench 🇹🇬 ### The Reference Benchmark for Evaluating LLMs in Ewe Language *Le benchmark de référence pour l'évaluation de LLMs en langue Ewe* [![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-orange.svg)](LICENSE) [![Tests: 107](https://img.shields.io/badge/Tests-107-blue.svg)](#categories) [![Categories: 10](https://img.shields.io/badge/Categories-10-green.svg)](#categories) [![Version: 1.0](https://img.shields.io/badge/Version-1.0-purple.svg)](CHANGELOG.md) [English](#english) • [Français](#français) • [Documentation](docs/) • [Leaderboard](#leaderboard) ---
## English ### What is ÈwéBench? ÈwéBench is the **first standardized benchmark** for evaluating Large Language Models (LLMs) on the **Ewe language** (ɛʋɛgbɛ) a Kwa language spoken by ~7 million people in Togo and Ghana. Unlike generic multilingual benchmarks that treat African languages as afterthoughts, ÈwéBench is **designed from the ground up** for Ewe, with culturally relevant tests, native speaker validation, and evaluation criteria that understand Ewe's unique linguistic features (tonality, agglutination, proverbs). ### Why ÈwéBench? - **No existing benchmark** specifically evaluates LLM capabilities in Ewe - Generic multilingual benchmarks (MMLU, HellaSwag) don't capture Ewe's nuances - African languages need **dedicated evaluation tools** to track real progress - Researchers and developers need a **common standard** to compare models ### Key Features | Feature | Description | |---------|-------------| | **10 categories** | From linguistic comprehension to agentic capabilities | | **107 tests** | Manually crafted, culturally grounded | | **Weighted scoring** | ÈwéScore single metric, weighted by category importance | | **Any model** | Works with any OpenAI-compatible API (local or cloud) | | **CLI & API** | Run from terminal or integrate into CI/CD | | **Leaderboard** | Track and compare model progress | | **Presets** | One-command evaluation for Gemini, local models | ### Quick Start ```bash # Clone the repo git clone https://github.com/joel710/EweBench.git cd EweBench # Install dependencies pip install -r requirements.txt # Run with a preset python run_benchmark.py --preset model --verbose # Run with a custom endpoint python run_benchmark.py --endpoint http://localhost:11434/v1/chat/completions \ --model yawo-v10 --verbose # Compare two results python run_benchmark.py --compare results/model_a.json results/model_b.json # View leaderboard python run_benchmark.py --leaderboard ``` ### Categories | # | Category | Tests | Weight | Description | |---|----------|-------|--------|-------------| | 1 | Linguistic Comprehension | 15 | 15% | Grammar, vocabulary, tonality, morphology | | 2 | Text Generation | 12 | 15% | Fluency, coherence, natural Ewe output | | 3 | Reasoning | 12 | 12% | Logical reasoning expressed in Ewe | | 4 | Translation | 12 | 12% | Bidirectional FR↔Ewe, EN↔Ewe | | 5 | Cultural Knowledge | 10 | 10% | Proverbs, traditions, Ewe/Togolese history | | 6 | Instruction Following | 10 | 10% | Complex instruction compliance | | 7 | Multi-turn | 8 | 8% | Context coherence across turns | | 8 | Agentic | 10 | 8% | Function calling, tool use | | 9 | Style Adaptation | 8 | 5% | Register switching (formal/informal) | | 10 | Robustness | 10 | 5% | Consistency under adversarial inputs | | | **Total** | **107** | **100%** | | ### Scoring ÈwéScore The **ÈwéScore** is a single number (0-100) representing overall Ewe language capability: ``` ÈwéScore = Σ (category_score × category_weight) / Σ active_weights ``` Each test is scored 0.0-1.0 using evaluation methods: - **exact_match** Normalized string comparison - **keywords** Presence of expected Ewe keywords - **multiple_choice** QCM answer detection - **format** Output format compliance (markdown, function_call, etc.) - **ewe_quality** Heuristic Ewe linguistic quality (character usage, vocabulary, structure) - **composite** Weighted combination of multiple methods **Passing threshold**: A test is "passed" if score ≥ 0.7 ### Evaluation Methods | Method | Use case | How it works | |--------|----------|--------------| | `exact_match` | Factual QA | Normalized comparison with expected answer | | `keywords` | Open-ended | Checks presence of expected Ewe keywords in response | | `multiple_choice` | QCM | Detects correct answer letter (A/B/C/D) | | `format` | Structured output | Validates format (markdown, function_call, length) | | `ewe_quality` | Free generation | Scores Ewe character usage, vocabulary, sentence structure | | `composite` | Complex tests | Average of keywords + ewe_quality + format | ### API Compatibility ÈwéBench works with any API implementing the OpenAI chat completions format: ``` POST /v1/chat/completions { "model": "model-name", "messages": [{"role": "user", "content": "..."}], "temperature": 0.3, "max_tokens": 1024 } ``` **Tested providers:** - Openai SDK - Google Gemini (OpenAI-compatible endpoint) - Ollama (local) - vLLM (local) - Any OpenAI-compatible server --- ## Français ### Qu'est-ce qu'ÈwéBench ? ÈwéBench est le **premier benchmark standardisé** pour évaluer les grands modèles de langage (LLMs) sur la **langue Ewe** (ɛʋɛgbɛ) une langue Kwa parlée par ~7 millions de personnes au Togo et au Ghana. Contrairement aux benchmarks multilingues génériques qui traitent les langues africaines comme des détails, ÈwéBench est **conçu de zéro** pour l'Ewe, avec des tests culturellement pertinents, une validation par des locuteurs natifs, et des critères d'évaluation qui comprennent les particularités linguistiques de l'Ewe (tonalité, agglutination, proverbes). ### Pourquoi ÈwéBench ? - **Aucun benchmark existant** n'évalue spécifiquement les capacités LLM en Ewe - Les benchmarks multilingues génériques (MMLU, HellaSwag) ne capturent pas les nuances de l'Ewe - Les langues africaines ont besoin d'**outils d'évaluation dédiés** pour mesurer les vrais progrès - Les chercheurs et développeurs ont besoin d'un **standard commun** pour comparer les modèles ### Démarrage rapide ```bash # Cloner le repo git clone https://github.com/joel710/EweBench.git cd EweBench # Installer les dépendances pip install -r requirements.txt # Configurer (optionnel pour les presets cloud) cp .env.example .env # Ajouter vos clés API dans .env # Lancer avec un preset python run_benchmark.py --preset model --verbose # Lancer sur un modèle local python run_benchmark.py --endpoint http://localhost:11434/v1/chat/completions \ --model yawo-v10 --verbose # Évaluer une seule catégorie python run_benchmark.py --preset model --category cultural_knowledge -v # Comparer deux modèles python run_benchmark.py --compare results/model.json results/yawo.json # Voir le classement python run_benchmark.py --leaderboard ``` ### Catégories | # | Catégorie | Tests | Poids | Description | |---|-----------|-------|-------|-------------| | 1 | Compréhension Linguistique | 15 | 15% | Grammaire, vocabulaire, tons, morphologie | | 2 | Génération de Texte | 12 | 15% | Fluence, cohérence, naturel du texte Ewe | | 3 | Raisonnement | 12 | 12% | Raisonnement logique exprimé en Ewe | | 4 | Traduction | 12 | 12% | Bidirectionnelle FR↔Ewe, EN↔Ewe | | 5 | Connaissance Culturelle | 10 | 10% | Proverbes, traditions, histoire Ewe/togolaise | | 6 | Suivi d'Instructions | 10 | 10% | Respect d'instructions complexes | | 7 | Multi-tour | 8 | 8% | Cohérence contextuelle sur plusieurs échanges | | 8 | Agentique | 10 | 8% | Function calling, utilisation d'outils | | 9 | Adaptation Stylistique | 8 | 5% | Registres formel/informel, technique/simple | | 10 | Robustesse | 10 | 5% | Cohérence face aux entrées adverses | | | **Total** | **107** | **100%** | | ### Scoring ÈwéScore L'**ÈwéScore** est un nombre unique (0-100) représentant la capacité globale en Ewe : ``` ÈwéScore = Σ (score_catégorie × poids_catégorie) / Σ poids_actifs ``` **Seuil de réussite** : Un test est "réussi" si le score ≥ 0.7 --- ## Leaderboard | # | Model | ÈwéScore | Tests Passed | Date | |---|-------|----------|--------------|------| | 🥇 | *En attente de soumissions* | - | - | - | > **Soumettre vos résultats** : Exécutez le benchmark, puis ouvrez une PR avec votre fichier de résultats dans `results/`. --- ## Project Structure ``` EweBench/ ├── README.md # This file (bilingual EN/FR) ├── LICENSE # CC BY-NC 4.0 ├── requirements.txt # Python dependencies ├── .env.example # API keys template ├── ewe_bench.py # Core benchmark engine ├── run_benchmark.py # CLI runner with presets ├── leaderboard.json # Public leaderboard data ├── tests/ # Test suites (107 tests) │ ├── linguistic_comprehension.json (15) │ ├── text_generation.json (12) │ ├── reasoning.json (12) │ ├── translation.json (12) │ ├── cultural_knowledge.json (10) │ ├── instruction_following.json (10) │ ├── multi_turn.json (8) │ ├── agentic.json (10) │ ├── style_adaptation.json (8) │ └── robustness.json (10) ├── results/ # Benchmark results (gitignored) ├── docs/ │ ├── METHODOLOGY.md # Scoring methodology details │ ├── CONTRIBUTING.md # How to contribute tests │ └── TEST_FORMAT.md # Test JSON format specification └── .github/ └── ISSUE_TEMPLATE.md ``` --- ## Contributing We welcome contributions! See [docs/CONTRIBUTING.md](docs/CONTRIBUTING.md) for details. Ways to contribute: - **Add tests** More tests improve coverage - **Validate translations** Native speaker review - **Submit results** Run on your model and share - **Report issues** Found a bad test? Let us know --- ## License **CC BY-NC 4.0** Creative Commons Attribution-NonCommercial 4.0 International - ✅ Free to use for research, education, and evaluation - ✅ Free to modify and redistribute (with attribution) - ⚠️ Commercial use requires explicit permission from Joel Elisée ADZONYA / Strive AI --- ## Citation ```bibtex @misc{ewebench2026, author = {Joel Elisée ADZONYA}, title = {ÈwéBench: A Reference Benchmark for Evaluating LLMs in Ewe Language}, year = {2026}, publisher = {Strive AI}, howpublished = {\url{https://github.com/joel710/EweBench}} } ``` ---
**Created by [Joel Elisée ADZONYA](https://joel.adzonya.strivenew.com) [Strive AI](https://github.com/joel710)** *L'IA au service des langues africaines*