docs/METHODOLOGY.md

ÈwéBench — Scoring Methodology

English • Français

---

English

Overview

ÈwéBench uses a weighted multi-category scoring system. Each model receives a single ÈwéScore (0-100) that represents its overall capability in Ewe.

Formula

ÈwéScore = Σ (category_score × category_weight) / Σ active_weights

Where:

• category_score = average test score within a category (0-100)
• category_weight = importance weight of that category
• active_weights = sum of weights for categories that were actually evaluated (handles partial runs)

Weight Rationale

Category	Weight	Justification
Linguistic Comprehension	15%	Core: understanding Ewe grammar, tones, morphology
Text Generation	15%	Core: producing natural, fluent Ewe text
Reasoning	12%	Important: expressing logical thought in Ewe
Translation	12%	Important: practical bilingual capability
Cultural Knowledge	10%	Valuable: proverbs, traditions, history
Instruction Following	10%	Practical: real-world usability
Multi-turn	8%	Advanced: conversation coherence
Agentic	8%	Advanced: tool use and planning
Style Adaptation	5%	Bonus: register switching
Robustness	5%	Bonus: adversarial resilience

Weights sum to 100%. Categories are ordered by importance: language mastery first, then practical capabilities, then advanced features.

Test Scoring

Each individual test is scored 0.0 to 1.0 using one of these methods:

1. Exact Match (exact_match)

score = 1.0 if normalize(expected) == normalize(response) else 0.0

Used for factual questions with a single correct answer.

2. Keyword Presence (keywords)

score = count(found_keywords) / count(expected_keywords)

Used for open-ended questions where specific Ewe terms should appear.

3. Multiple Choice (multiple_choice)

score = 1.0 if correct_letter detected in response else 0.0

Used for QCM-style tests with A/B/C/D options.

4. Format Compliance (format)

Checks multiple format criteria:

• contains_ewe — Response has Ewe characters (ɖ, ɛ, ɔ, ƒ, ŋ, ɣ)
• min_length / max_length — Response length bounds
• contains_function_call — Has <function_call> tag
• markdown_elements — Has tables, headers, lists, bold

5. Ewe Quality Heuristic (ewe_quality)

Composite heuristic scoring:

• +0.3 for Ewe special characters presence
• +0.05 per common Ewe word found (max +0.4)
• -0.2 if too many French words detected (>5)
• +0.2 for multi-sentence structure
• +0.1 for minimum response length

6. Composite (composite)

score = (keywords_score + ewe_quality_score + format_score) / 3

Used for complex tests requiring multiple evaluation dimensions.

Pass/Fail Threshold

A test is passed if score >= 0.7.

This threshold balances:

• Not too strict (some Ewe variability is expected)
• Not too lenient (ensures meaningful output quality)

Category Score

category_score = (sum of test scores / number of tests) × 100

---

Français

Vue d'ensemble

ÈwéBench utilise un système de scoring multi-catégories pondéré. Chaque modèle reçoit un ÈwéScore unique (0-100) représentant sa capacité globale en Ewe.

Formule

ÈwéScore = Σ (score_catégorie × poids_catégorie) / Σ poids_actifs

Justification des poids

Catégorie	Poids	Justification
Compréhension Linguistique	15%	Cœur : compréhension grammaire, tons, morphologie Ewe
Génération de Texte	15%	Cœur : production de texte Ewe naturel et fluide
Raisonnement	12%	Important : expression de la pensée logique en Ewe
Traduction	12%	Important : capacité bilingue pratique
Connaissance Culturelle	10%	Précieux : proverbes, traditions, histoire
Suivi d'Instructions	10%	Pratique : utilisabilité réelle
Multi-tour	8%	Avancé : cohérence conversationnelle
Agentique	8%	Avancé : utilisation d'outils et planification
Adaptation Stylistique	5%	Bonus : changement de registre
Robustesse	5%	Bonus : résilience adversariale

Seuil de réussite

Un test est réussi si score >= 0.7.

Score par catégorie

score_catégorie = (somme des scores de tests / nombre de tests) × 100

---

Known Limitations

1. Ewe quality heuristic is rule-based, not learned — it can miss valid Ewe or reward superficial patterns
2. Keyword matching doesn't account for synonyms or paraphrasing
3. No human evaluation in automated runs — ÈwéScore is an approximation
4. Tonal accuracy cannot be verified in written text (Ewe is tonal but rarely written with tone marks)

These limitations are documented so users interpret scores with appropriate context. We plan to add human evaluation protocols in v2.0.