The AI Testing Cheat Sheet Nobody Told You About

Measures overlap between generated text and reference text using n-gram precision. Range: 0-1 (higher is better). Penalizes length mismatches.

Use when: Evaluating machine translation, summarization, or any task with fixed reference outputs.

Don't use when: There are multiple valid correct answers. BLEU hates paraphrasing.

Variant of BLEU that emphasizes recall instead of precision. ROUGE-1 (unigrams), ROUGE-2 (bigrams), ROUGE-L (longest common subsequence).

Use when: Evaluating summarization. ROUGE-L is particularly good for factual consistency.

Don't use when: You need semantic understanding. ROUGE is pure overlap-based.

Uses contextual embeddings from BERT to measure semantic similarity between generated and reference text. Computes precision, recall, and F1. Range: 0-1.

Use when: You care about meaning, not surface-level n-gram overlap. Works well for summarization and paraphrase detection.

Don't use when: You need to know if specific entities were preserved (use NER-based metrics instead).

Aligns words using exact match, stem match, and synonymy. Penalizes reordering. Often correlates better with human judgment than BLEU.

Use when: Translating, paraphrasing, or anything where word reordering is acceptable but penalizable.

Don't use when: Speed matters (METEOR is slow due to synonym lookup).

Embeds both generated and reference text, computes cosine distance between vectors. Range: -1 to 1 (higher is better). Fast, lightweight.

Use when: You need a quick semantic check. Real-time monitoring. Continuous evaluation.

Don't use when: You need fine-grained semantic analysis. Embeddings lose detail.

Measures whether generated output is consistent with source documents. Typically binary (consistent/inconsistent) or scaled (0-1). Can use NLI models, semantic similarity, or question generation.

Use when: Your LLM must not hallucinate. Critical for RAG, summarization, question-answering.

Don't use when: You're generating creative content (poetry, fiction).

Uses a model trained on entailment tasks to check if generated text is logically consistent with reference text. Scores entailment, contradiction, or neutral.

Use when: You need semantic-level fact-checking. Works across domains without retraining.

Don't use when: You need domain-specific fact verification (medical claims, legal citations).

Measures how much of the generated text is directly copied or closely paraphrased from source documents. Simple but effective.

Use when: You want to know if the model is grounded in your sources (not hallucinating).

Don't use when: You need to measure abstractive quality (this penalizes good summaries).

Systematically tests model behavior across predefined scenarios: negation, argument reversal, numeric variation, etc. Framework popularized by Ribeiro et al. (2020).

Use when: You want Practical coverage of edge cases. Catching adversarial inputs. Stress testing.

Don't use when: You don't have domain-specific scenarios (hard to write good checks).

Tests how the model behaves under intentional perturbations: typos, paraphrasing, noise injection, or semantic attacks.

Use when: You're concerned about adversarial inputs or distribution shift in production.

Don't use when: Your deployment is controlled (enterprise internal tool).

Measures how much the output changes with minor prompt variations. High sensitivity = fragile model. Low sensitivity = robust.

Use when: You're building user-facing systems (users will vary their inputs).

Don't use when: Your prompts are fixed and controlled.

Monitors whether production input distributions differ significantly from training data. Uses statistical tests (KS test, MMD) or distributional distance metrics.

Use when: You're monitoring production models continuously.

Don't use when: Your inputs are known to vary (intentional domain shift).

Tracks changes in evaluation metrics over time. Alerts when performance drops below baseline. Rolling window comparison.

Use when: You need early warning of model performance issues.

Don't use when: You're okay with discovering problems reactively.

Evaluates confidence in human-provided labels. Flags low-quality labels, disagreement patterns, and outliers.

Use when: You're building labeled datasets (especially crowdsourced).

Don't use when: Your labels come from single expert (already high confidence).

Binary metric: generated answer matches reference answer exactly (case-sensitive or case-insensitive). Range: 0-1.

Use when: The answer must be precise. QA tasks with single correct answer.

Don't use when: Multiple answers are valid (use F1 instead).

Harmonic mean of precision and recall. Ranges 0-1. Good for comparing overlapping answers.

Use when: Multiple valid answers exist or partial credit is meaningful (extractive QA).

Don't use when: You need binary success/failure (use EM).

For ranking tasks. Measures position of first correct item. Rewards putting correct answers high.

Use when: Evaluating search, retrieval, or ranking systems.

Don't use when: You don't care about ranking order (use recall instead).

Measures ranking quality. Discounts items lower in the ranking. Normalized to 0-1. More sophisticated than MRR.

Use when: Evaluating recommendation systems or search results where position matters.

Don't use when: You just need simple recall metrics.

Percentage of outputs containing factually incorrect information not in the source. Measured manually or with NLI models.

Use when: You need to track how often your model makes things up.

Don't use when: You're evaluating creative tasks (hallucination is expected).

Measures whether model's stated confidence matches actual correctness. Well-calibrated model: 90% confidence = 90% correct.

Use when: You're using model confidence for decisions (rejection thresholds, routing).

Don't use when: You're not using confidence scores operationally.

Examines entropy or probability mass of generated tokens. High entropy = uncertain. Low entropy = confident.

Use when: You want fine-grained understanding of model uncertainty.

Don't use when: You just need pass/fail metrics.