Model Evaluation Metrics Explainer
Quick reference for ML and LLM evaluation metrics - BLEU, ROUGE, F1, pass@k, BERTScore and more - with formulas, ranges and use cases, plus an interactive confusion matrix calculator.
FreeOnline Tool
Loading…
How to Use
- Click a task tab (Text Generation, Classification, Translation, QA, Code Generation, or All) to filter the metrics for that task.
- Browse the metric cards: each shows the name and aliases, formula, value range, plain-language explanation and applicable task tags.
- For classification metrics, enter TP, FP, FN and TN in the confusion matrix calculator.
- Accuracy, precision, recall and F1 are computed in real time as percentages.
- Click 'Load Sample Data' to restore the default example values.
Features
- Covers 13 common metrics: BLEU, ROUGE, METEOR, Perplexity, Accuracy, Precision, Recall, F1, Exact Match, pass@k, BERTScore, ELO and ROUGE-L.
- Every card includes the formula, value range, plain-language explanation and applicable task tags.
- Filter by task type: text generation, classification, translation, QA and code generation.
- Built-in interactive confusion matrix calculator for accuracy, precision, recall and F1.
- Spans classical NLP metrics and LLM-era metrics (pass@k, ELO, BERTScore), with notes on when each metric is appropriate and its limitations.
Use Cases
Look up a metric while reading
Came across BLEU-4 or pass@10 in a paper? Open the card to see the formula and what it really measures.
Design an evaluation plan
Filter by task type to pick suitable metrics, e.g. BLEU/METEOR for translation, Exact Match for QA and pass@k for code generation.
Compute classification results
Enter TP/FP/FN/TN from your classifier and instantly get accuracy, precision, recall and F1.
Learn evaluation concepts
Read through the cards by task to understand the difference between n-gram and semantic metrics, or how ELO rankings work.
FAQ
Is this data up to date?
Metric definitions and formulas are standard and stable knowledge. The tool covers both classical metrics and LLM-era ones (pass@k, ELO, BERTScore) for reference.
Can Perplexity be compared across models?
No. Perplexity from models with different tokenizers is not directly comparable - it is only meaningful under the same tokenization scheme, as noted in the tool.
Does the confusion matrix accept decimals?
Yes, the four inputs are parsed as numbers and results are computed in real time, displayed as percentages.
Is F-beta covered?
The F1 card mentions F-beta as a generalized form where beta greater than 1 emphasizes recall; see the card for details.
What is the ELO rating?
A ranking system from chess, adopted by LMSYS Chatbot Arena for LLM comparison. It updates scores based on pairwise wins between two models responses to the same prompt.