# Define success criteria and build evaluations Define measurable success criteria for your LLM application and build evaluations to test it, from exact match checks to LLM-based grading. --- Building a successful LLM-based application starts with clearly defining your success criteria and then designing evaluations to measure performance against them. This cycle is central to prompt engineering. ![Flowchart of prompt engineering: test cases, preliminary prompt, iterative testing and refinement, final validation, ship](/docs/images/how-to-prompt-eng.png) ## Define your success criteria Good success criteria are: * **Specific:** Clearly define what you want to achieve. Instead of "good performance," specify "accurate sentiment classification." * **Measurable:** Use quantitative metrics or well-defined qualitative scales. Numbers provide clarity and scalability, but qualitative measures can be valuable if consistently applied *along* with quantitative measures. * Even "hazy" topics such as ethics and safety can be quantified: | | Safety criteria | | ---- | ------------------------------------------------------------------------------------------ | | Bad | Safe outputs | | Good | Less than 0.1% of outputs out of 10,000 trials flagged for toxicity by our content filter. | **Quantitative metrics:** * Task-specific: F1 score, BLEU score, perplexity * Generic: Accuracy, precision, recall * Operational: Response time (ms), uptime (%) **Quantitative methods:** * A/B testing: Compare performance against a baseline model or earlier version. * User feedback: Implicit measures like task completion rates. * Edge case analysis: Percentage of edge cases handled without errors. **Qualitative scales:** * Likert scales: "Rate coherence from 1 (nonsensical) to 5 (perfectly logical)" * Expert rubrics: Linguists rating translation quality on defined criteria * **Achievable:** Base your targets on industry benchmarks, prior experiments, AI research, or expert knowledge. Your success metrics should not be unrealistic to current frontier model capabilities. * **Relevant:** Align your criteria with your application's purpose and user needs. Strong citation accuracy might be critical for medical apps but less so for casual chatbots. | | Criteria | | ---- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Bad | The model should classify sentiments well | | Good | Our sentiment analysis model should achieve an F1 score of at least 0.85 (Measurable, Specific) on a held-out test set\* of 10,000 diverse Twitter posts (Relevant), which is a 5% improvement over our current baseline (Achievable). | \*More on held-out test sets in the next section. ### Common success criteria Here are some criteria that might be important for your use case. This list is non-exhaustive. How well does the model need to perform on the task? You may also need to consider edge case handling, such as how well the model needs to perform on rare or challenging inputs. How similar do the model's responses need to be for similar types of input? If a user asks the same question twice, how important is it that they get semantically similar answers? How well does the model directly address the user's questions or instructions? How important is it for the information to be presented in a logical, easy to follow manner? How well does the model's output style match expectations? How appropriate is its language for the target audience? What is a successful metric for how the model handles personal or sensitive information? Can it follow instructions not to use or share certain details? How effectively does the model use provided context? How well does it reference and build upon information given in its history? What is the acceptable response time for the model? This depends on your application's real-time requirements and user expectations. What is your budget for running the model? Consider factors like the cost for each API call, the size of the model, and the frequency of usage. Most use cases need multidimensional evaluation along several success criteria. | | Criteria | | ---- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | Bad | The model should classify sentiments well | | Good | On a held-out test set of 10,000 diverse Twitter posts, the sentiment analysis model should achieve: - an F1 score of at least 0.85 - 99.5% of outputs are non-toxic - 90% of errors would cause inconvenience, not egregious error\* - 95% response time \< 200ms | \*In reality, you would also define what "inconvenience" and "egregious" mean. *** ## Build evaluations ### Eval design principles 1. **Be task-specific:** Design evals that mirror your real-world task distribution. Don't forget to factor in edge cases! * Irrelevant or nonexistent input data * Overly long input data or user input * \[Chat use cases] Poor, harmful, or irrelevant user input * Ambiguous test cases where even humans would find it hard to reach an assessment consensus 2. **Automate when possible:** Structure questions to allow for automated grading (for example, multiple-choice, string match, code-graded, LLM-graded). 3. **Prioritize volume over quality:** More questions with slightly lower signal automated grading is better than fewer questions with high-quality human hand-graded evals. ### Example evals **What it measures:** Exact match evals measure whether the model's output matches a predefined correct answer, typically after normalizing whitespace and case. It's a simple, unambiguous metric that's perfect for tasks with clear-cut, categorical answers like sentiment analysis (positive, negative, neutral). **Example eval test cases:** 1,000 tweets with human-labeled sentiments. ```python Python tweets = [ {"text": "This movie was a total waste of time. 👎", "sentiment": "negative"}, {"text": "The new album is 🔥! Been on repeat all day.", "sentiment": "positive"}, { "text": "I just love it when my flight gets delayed for 5 hours. #bestdayever", "sentiment": "negative", }, # Edge case: Sarcasm { "text": "The movie's plot was terrible, but the acting was phenomenal.", "sentiment": "mixed", }, # Edge case: Mixed sentiment # ... 996 more tweets ] client = anthropic.Anthropic() def get_completion(prompt: str): message = client.messages.create( model="claude-opus-5", max_tokens=50, messages=[{"role": "user", "content": prompt}], ) return next(block.text for block in message.content if block.type == "text") def evaluate_exact_match(model_output, correct_answer): return model_output.strip().lower() == correct_answer.lower() outputs = [ get_completion( f"Classify this as 'positive', 'negative', 'neutral', or 'mixed': {tweet['text']}" ) for tweet in tweets ] accuracy = sum( evaluate_exact_match(output, tweet["sentiment"]) for output, tweet in zip(outputs, tweets) ) / len(tweets) print(f"Sentiment Analysis Accuracy: {accuracy * 100}%") ``` ```typescript TypeScript const tweets = [ { text: "This movie was a total waste of time. 👎", sentiment: "negative" }, { text: "The new album is 🔥! Been on repeat all day.", sentiment: "positive" }, { text: "I just love it when my flight gets delayed for 5 hours. #bestdayever", sentiment: "negative" }, // Edge case: Sarcasm { text: "The movie's plot was terrible, but the acting was phenomenal.", sentiment: "mixed" } // Edge case: Mixed sentiment // ... 996 more tweets ]; const client = new Anthropic(); async function getCompletion(prompt: string): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 50, messages: [{ role: "user", content: prompt }] }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } function evaluateExactMatch(modelOutput: string, correctAnswer: string): boolean { return modelOutput.trim().toLowerCase() === correctAnswer.toLowerCase(); } let correctCount = 0; for (const tweet of tweets) { const output = await getCompletion( `Classify this as 'positive', 'negative', 'neutral', or 'mixed': ${tweet.text}` ); if (evaluateExactMatch(output, tweet.sentiment)) { correctCount++; } } console.log(`Sentiment Analysis Accuracy: ${(correctCount / tweets.length) * 100}%`); ``` ```csharp C# Tweet[] tweets = [ new("This movie was a total waste of time. 👎", "negative"), new("The new album is 🔥! Been on repeat all day.", "positive"), // Edge case: Sarcasm new("I just love it when my flight gets delayed for 5 hours. #bestdayever", "negative"), // Edge case: Mixed sentiment new("The movie's plot was terrible, but the acting was phenomenal.", "mixed"), // ... 996 more tweets ]; var client = new AnthropicClient(); async Task GetCompletion(string prompt) { var message = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 50, Messages = [new() { Role = Role.User, Content = prompt }], }); return ContentText(message); } bool EvaluateExactMatch(string modelOutput, string correctAnswer) { return string.Equals(modelOutput.Trim(), correctAnswer, StringComparison.OrdinalIgnoreCase); } string ContentText(Message message) { var text = ""; foreach (var block in message.Content) { if (block.TryPickText(out var textBlock)) { text += textBlock.Text; } } return text; } var correct = 0; foreach (var tweet in tweets) { var output = await GetCompletion( $"Classify this as 'positive', 'negative', 'neutral', or 'mixed': {tweet.Text}"); if (EvaluateExactMatch(output, tweet.Sentiment)) { correct++; } } Console.WriteLine($"Sentiment Analysis Accuracy: {100.0 * correct / tweets.Length}%"); record Tweet(string Text, string Sentiment); ``` ```go Go var client = anthropic.NewClient() func contentText(message *anthropic.Message) string { var text strings.Builder for _, block := range message.Content { if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok { text.WriteString(textBlock.Text) } } return text.String() } type tweet struct { Text string Sentiment string } var tweets = []tweet{ {"This movie was a total waste of time. 👎", "negative"}, {"The new album is 🔥! Been on repeat all day.", "positive"}, // Edge case: Sarcasm {"I just love it when my flight gets delayed for 5 hours. #bestdayever", "negative"}, // Edge case: Mixed sentiment {"The movie's plot was terrible, but the acting was phenomenal.", "mixed"}, // ... 996 more tweets } func getCompletion(prompt string) string { message, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 50, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(prompt)), }, }) if err != nil { log.Fatal(err) } return contentText(message) } func evaluateExactMatch(modelOutput, correctAnswer string) bool { return strings.EqualFold(strings.TrimSpace(modelOutput), correctAnswer) } func main() { correct := 0 for _, item := range tweets { output := getCompletion("Classify this as 'positive', 'negative', 'neutral', or 'mixed': " + item.Text) if evaluateExactMatch(output, item.Sentiment) { correct++ } } fmt.Printf("Sentiment Analysis Accuracy: %.1f%%\n", float64(correct)/float64(len(tweets))*100) } ``` ```java Java record Tweet(String text, String sentiment) {} List tweets = List.of( new Tweet("This movie was a total waste of time. 👎", "negative"), new Tweet("The new album is 🔥! Been on repeat all day.", "positive"), // Edge case: Sarcasm new Tweet("I just love it when my flight gets delayed for 5 hours. #bestdayever", "negative"), // Edge case: Mixed sentiment new Tweet("The movie's plot was terrible, but the acting was phenomenal.", "mixed") // ... 996 more tweets ); AnthropicClient client = AnthropicOkHttpClient.fromEnv(); String contentText(Message message) { var text = new StringBuilder(); for (var block : message.content()) { block.text().ifPresent(textBlock -> text.append(textBlock.text())); } return text.toString(); } String getCompletion(String prompt) { var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(50L) .addUserMessage(prompt) .build(); return contentText(client.messages().create(params)); } boolean evaluateExactMatch(String modelOutput, String correctAnswer) { return modelOutput.strip().equalsIgnoreCase(correctAnswer); } void main() { int correct = 0; for (var tweet : tweets) { var output = getCompletion( "Classify this as 'positive', 'negative', 'neutral', or 'mixed': " + tweet.text()); if (evaluateExactMatch(output, tweet.sentiment())) { correct++; } } IO.println("Sentiment Analysis Accuracy: " + (100.0 * correct / tweets.size()) + "%"); } ``` ```php PHP $client = new Client(); $tweets = [ ['text' => 'This movie was a total waste of time. 👎', 'sentiment' => 'negative'], ['text' => 'The new album is 🔥! Been on repeat all day.', 'sentiment' => 'positive'], // Edge case: Sarcasm ['text' => 'I just love it when my flight gets delayed for 5 hours. #bestdayever', 'sentiment' => 'negative'], // Edge case: Mixed sentiment ['text' => "The movie's plot was terrible, but the acting was phenomenal.", 'sentiment' => 'mixed'], // ... 996 more tweets ]; function getCompletion(Client $client, string $prompt): string { $message = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 50, messages: [ [ 'role' => 'user', 'content' => $prompt, ], ], ); return contentText($message); } function evaluateExactMatch(string $modelOutput, string $correctAnswer): bool { return strtolower(trim($modelOutput)) === strtolower($correctAnswer); } function contentText($message): string { $text = ''; foreach ($message->content as $block) { if ($block instanceof TextBlock) { $text .= $block->text; } } return $text; } $correct = 0; foreach ($tweets as $tweet) { $output = getCompletion( $client, "Classify this as 'positive', 'negative', 'neutral', or 'mixed': {$tweet['text']}", ); if (evaluateExactMatch($output, $tweet['sentiment'])) { $correct++; } } echo 'Sentiment Analysis Accuracy: ' . (100 * $correct / count($tweets)) . '%' . PHP_EOL; ``` ```ruby Ruby client = Anthropic::Client.new tweets = [ { text: "This movie was a total waste of time. 👎", sentiment: "negative" }, { text: "The new album is 🔥! Been on repeat all day.", sentiment: "positive" }, # Edge case: Sarcasm { text: "I just love it when my flight gets delayed for 5 hours. #bestdayever", sentiment: "negative" }, # Edge case: Mixed sentiment { text: "The movie's plot was terrible, but the acting was phenomenal.", sentiment: "mixed" } # ... 996 more tweets ] def content_text(message) message.content.filter_map { |block| block.text if block.type == :text }.join end def get_completion(client, prompt) message = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 50, messages: [ { role: "user", content: prompt } ] ) content_text(message) end def evaluate_exact_match(model_output, correct_answer) model_output.strip.downcase == correct_answer.downcase end correct = tweets.count do |tweet| output = get_completion( client, "Classify this as 'positive', 'negative', 'neutral', or 'mixed': #{tweet[:text]}" ) evaluate_exact_match(output, tweet[:sentiment]) end puts "Sentiment Analysis Accuracy: #{100.0 * correct / tweets.length}%" ``` **What it measures:** Cosine similarity measures the similarity between two vectors (in this case, sentence embeddings of the model's output using [Sentence-BERT (SBERT)](https://sbert.net/)) by computing the cosine of the angle between them. Values closer to 1 indicate higher similarity. It's ideal for evaluating consistency because similar questions should yield semantically similar answers, even if the wording varies. **Example eval test cases:** 50 groups with a few paraphrased versions each. ```python Python from sentence_transformers import SentenceTransformer import numpy as np # ... faq_variations = [ { "questions": [ "What's your return policy?", "How can I return an item?", "Wut's yur retrn polcy?", ], "answer": "Our return policy allows...", }, # Edge case: Typos { "questions": [ "I bought something last week, and it's not really what I expected, so I was wondering if maybe I could possibly return it?", "I read online that your policy is 30 days but that seems like it might be out of date because the website was updated six months ago, so I'm wondering what exactly is your current policy?", ], "answer": "Our return policy allows...", }, # Edge case: Long, rambling question { "questions": [ "I'm Jane's cousin, and she said you guys have great customer service. Can I return this?", "Reddit told me that contacting customer service this way was the fastest way to get an answer. I hope they're right! What is the return window for a jacket?", ], "answer": "Our return policy allows...", }, # Edge case: Irrelevant info # ... 47 more FAQs ] client = anthropic.Anthropic() def get_completion(prompt: str): message = client.messages.create( model="claude-opus-5", max_tokens=2048, messages=[{"role": "user", "content": prompt}], ) return next(block.text for block in message.content if block.type == "text") def evaluate_cosine_similarity(outputs): model = SentenceTransformer("all-MiniLM-L6-v2") embeddings = model.encode(outputs) norms = np.linalg.norm(embeddings, axis=1) cosine_similarities = np.dot(embeddings, embeddings.T) / np.outer(norms, norms) return np.mean(cosine_similarities) for faq in faq_variations: outputs = [get_completion(question) for question in faq["questions"]] similarity_score = evaluate_cosine_similarity(outputs) print(f"FAQ Consistency Score: {similarity_score * 100}%") ``` ```typescript TypeScript import { pipeline } from "@huggingface/transformers"; const faqVariations = [ { questions: [ "What's your return policy?", "How can I return an item?", "Wut's yur retrn polcy?" ], answer: "Our return policy allows..." }, // Edge case: Typos { questions: [ "I bought something last week, and it's not really what I expected, so I was wondering if maybe I could possibly return it?", "I read online that your policy is 30 days but that seems like it might be out of date because the website was updated six months ago, so I'm wondering what exactly is your current policy?" ], answer: "Our return policy allows..." }, // Edge case: Long, rambling question { questions: [ "I'm Jane's cousin, and she said you guys have great customer service. Can I return this?", "Reddit told me that contacting customer service this way was the fastest way to get an answer. I hope they're right! What is the return window for a jacket?" ], answer: "Our return policy allows..." } // Edge case: Irrelevant info // ... 47 more FAQs ]; const client = new Anthropic(); async function getCompletion(prompt: string): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 2048, messages: [{ role: "user", content: prompt }] }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } async function evaluateCosineSimilarity(outputs: string[]): Promise { const extractor = await pipeline("feature-extraction", "Xenova/all-MiniLM-L6-v2"); const embeddings = (await extractor(outputs, { pooling: "mean", normalize: true })).tolist(); let total = 0; for (const embeddingA of embeddings) { for (const embeddingB of embeddings) { // Vectors are normalized, so cosine similarity is the dot product total += embeddingA.reduce( (sum: number, value: number, i: number) => sum + value * embeddingB[i], 0 ); } } return total / (embeddings.length * embeddings.length); } for (const faq of faqVariations) { const outputs: string[] = []; for (const question of faq.questions) { outputs.push(await getCompletion(question)); } const similarityScore = await evaluateCosineSimilarity(outputs); console.log(`FAQ Consistency Score: ${similarityScore * 100}%`); } ``` ```csharp C# // Sentence-embedding models are not available as a native C# library. See the Python or TypeScript tab for this eval recipe. ``` ```go Go // Sentence-embedding models are not available as a native Go library. See the Python or TypeScript tab for this eval recipe. ``` ```java Java // Sentence-embedding models are not available as a native Java library. See the Python or TypeScript tab for this eval recipe. ``` ```php PHP // Sentence-embedding models are not available as a native PHP library. See the Python or TypeScript tab for this eval recipe. ``` ```ruby Ruby # Sentence-embedding models are not available as a native Ruby library. See the Python or TypeScript tab for this eval recipe. ``` **What it measures:** ROUGE-L (Recall-Oriented Understudy for Gisting Evaluation - Longest Common Subsequence) evaluates the quality of generated summaries. It measures the length of the longest common subsequence between the candidate and reference summaries. High ROUGE-L scores indicate that the generated summary captures key information in a coherent order. **Example eval test cases:** 200 articles with reference summaries. ```python Python from rouge import Rouge # ... articles = [ { "text": "In a groundbreaking study, researchers at MIT...", "summary": "MIT scientists discover a new antibiotic...", }, { "text": "Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...", "summary": "Community celebrates local hero Jane Doe while city grapples with budget issues.", }, # Edge case: Multitopic { "text": "You won't believe what this celebrity did! ... extensive charity work ...", "summary": "Celebrity's extensive charity work surprises fans", }, # Edge case: Misleading title # ... 197 more articles ] client = anthropic.Anthropic() def get_completion(prompt: str): message = client.messages.create( model="claude-opus-5", max_tokens=1024, messages=[{"role": "user", "content": prompt}], ) return next(block.text for block in message.content if block.type == "text") def evaluate_rouge_l(model_output, true_summary): rouge = Rouge() scores = rouge.get_scores(model_output, true_summary) return scores[0]["rouge-l"]["f"] # ROUGE-L F1 score outputs = [ get_completion(f"Summarize this article in 1-2 sentences:\n\n{article['text']}") for article in articles ] relevance_scores = [ evaluate_rouge_l(output, article["summary"]) for output, article in zip(outputs, articles) ] print(f"Average ROUGE-L F1 Score: {sum(relevance_scores) / len(relevance_scores)}") ``` ```typescript TypeScript const articles = [ { text: "In a groundbreaking study, researchers at MIT...", summary: "MIT scientists discover a new antibiotic..." }, { text: "Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...", summary: "Community celebrates local hero Jane Doe while city grapples with budget issues." }, // Edge case: Multitopic { text: "You won't believe what this celebrity did! ... extensive charity work ...", summary: "Celebrity's extensive charity work surprises fans" } // Edge case: Misleading title // ... 197 more articles ]; const client = new Anthropic(); async function getCompletion(prompt: string): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 1024, messages: [{ role: "user", content: prompt }] }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } // ROUGE-L measures the longest common subsequence (LCS) of words between the // candidate and reference summaries, reported here as an F1 score. Tokenization // is simplified to whitespace words; scores may differ from the Python rouge library. function rougeL(candidate: string, reference: string): number { const candidateWords = candidate.toLowerCase().trim().split(/\s+/); const referenceWords = reference.toLowerCase().trim().split(/\s+/); const lcsLengths: number[][] = Array.from({ length: candidateWords.length + 1 }, () => new Array(referenceWords.length + 1).fill(0) ); for (const [i, candidateWord] of candidateWords.entries()) { for (const [j, referenceWord] of referenceWords.entries()) { lcsLengths[i + 1][j + 1] = candidateWord === referenceWord ? lcsLengths[i][j] + 1 : Math.max(lcsLengths[i][j + 1], lcsLengths[i + 1][j]); } } const lcs = lcsLengths[candidateWords.length][referenceWords.length]; if (lcs === 0) return 0; const precision = lcs / candidateWords.length; const recall = lcs / referenceWords.length; return (2 * precision * recall) / (precision + recall); } const relevanceScores: number[] = []; for (const article of articles) { const output = await getCompletion( `Summarize this article in 1-2 sentences:\n\n${article.text}` ); relevanceScores.push(rougeL(output, article.summary)); } const averageScore = relevanceScores.reduce((sum, score) => sum + score, 0) / relevanceScores.length; console.log(`Average ROUGE-L F1 Score: ${averageScore}`); ``` ```csharp C# using System.Text.RegularExpressions; // ... Article[] articles = [ new("In a groundbreaking study, researchers at MIT...", "MIT scientists discover a new antibiotic..."), // Edge case: Multitopic new("Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...", "Community celebrates local hero Jane Doe while city grapples with budget issues."), // Edge case: Misleading title new("You won't believe what this celebrity did! ... extensive charity work ...", "Celebrity's extensive charity work surprises fans"), // ... 197 more articles ]; var client = new AnthropicClient(); async Task GetCompletion(string prompt) { var message = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 1024, Messages = [new() { Role = Role.User, Content = prompt }], }); return ContentText(message); } // ROUGE-L measures the longest common subsequence (LCS) of words between the // candidate and reference summaries, reported here as an F1 score. Tokenization // is simplified to whitespace words; scores may differ from the Python rouge library. double RougeL(string candidate, string reference) { var candidateWords = Regex.Split(candidate.ToLowerInvariant().Trim(), @"\s+"); var referenceWords = Regex.Split(reference.ToLowerInvariant().Trim(), @"\s+"); var lcsLengths = new int[candidateWords.Length + 1, referenceWords.Length + 1]; for (var i = 0; i < candidateWords.Length; i++) { for (var j = 0; j < referenceWords.Length; j++) { lcsLengths[i + 1, j + 1] = candidateWords[i] == referenceWords[j] ? lcsLengths[i, j] + 1 : Math.Max(lcsLengths[i, j + 1], lcsLengths[i + 1, j]); } } var lcs = lcsLengths[candidateWords.Length, referenceWords.Length]; if (lcs == 0) { return 0; } var precision = (double)lcs / candidateWords.Length; var recall = (double)lcs / referenceWords.Length; return 2 * precision * recall / (precision + recall); } string ContentText(Message message) { var text = ""; foreach (var block in message.Content) { if (block.TryPickText(out var textBlock)) { text += textBlock.Text; } } return text; } var relevanceScores = new List(); foreach (var article in articles) { var output = await GetCompletion($"Summarize this article in 1-2 sentences:\n\n{article.Text}"); relevanceScores.Add(RougeL(output, article.Summary)); } Console.WriteLine($"Average ROUGE-L F1 Score: {relevanceScores.Average()}"); record Article(string Text, string Summary); ``` ```go Go var client = anthropic.NewClient() func contentText(message *anthropic.Message) string { var text strings.Builder for _, block := range message.Content { if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok { text.WriteString(textBlock.Text) } } return text.String() } type article struct { Text string Summary string } var articles = []article{ { "In a groundbreaking study, researchers at MIT...", "MIT scientists discover a new antibiotic...", }, // Edge case: Multitopic { "Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...", "Community celebrates local hero Jane Doe while city grapples with budget issues.", }, // Edge case: Misleading title { "You won't believe what this celebrity did! ... extensive charity work ...", "Celebrity's extensive charity work surprises fans", }, // ... 197 more articles } func getCompletion(prompt string) string { message, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 1024, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(prompt)), }, }) if err != nil { log.Fatal(err) } return contentText(message) } // ROUGE-L measures the longest common subsequence (LCS) of words between the // candidate and reference summaries, reported here as an F1 score. Tokenization // is simplified to whitespace words; scores may differ from the Python rouge library. func rougeL(candidate, reference string) float64 { candidateWords := strings.Fields(strings.ToLower(candidate)) referenceWords := strings.Fields(strings.ToLower(reference)) lcsLengths := make([][]int, len(candidateWords)+1) for i := range lcsLengths { lcsLengths[i] = make([]int, len(referenceWords)+1) } for i, candidateWord := range candidateWords { for j, referenceWord := range referenceWords { if candidateWord == referenceWord { lcsLengths[i+1][j+1] = lcsLengths[i][j] + 1 } else { lcsLengths[i+1][j+1] = max(lcsLengths[i][j+1], lcsLengths[i+1][j]) } } } lcs := lcsLengths[len(candidateWords)][len(referenceWords)] if lcs == 0 { return 0 } precision := float64(lcs) / float64(len(candidateWords)) recall := float64(lcs) / float64(len(referenceWords)) return 2 * precision * recall / (precision + recall) } func main() { var relevanceScores []float64 for _, item := range articles { output := getCompletion("Summarize this article in 1-2 sentences:\n\n" + item.Text) relevanceScores = append(relevanceScores, rougeL(output, item.Summary)) } total := 0.0 for _, score := range relevanceScores { total += score } fmt.Println("Average ROUGE-L F1 Score:", total/float64(len(relevanceScores))) } ``` ```java Java record Article(String text, String summary) {} List
articles = List.of( new Article( "In a groundbreaking study, researchers at MIT...", "MIT scientists discover a new antibiotic..."), // Edge case: Multitopic new Article( "Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...", "Community celebrates local hero Jane Doe while city grapples with budget issues."), // Edge case: Misleading title new Article( "You won't believe what this celebrity did! ... extensive charity work ...", "Celebrity's extensive charity work surprises fans") // ... 197 more articles ); AnthropicClient client = AnthropicOkHttpClient.fromEnv(); String contentText(Message message) { var text = new StringBuilder(); for (var block : message.content()) { block.text().ifPresent(textBlock -> text.append(textBlock.text())); } return text.toString(); } String getCompletion(String prompt) { var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(1024L) .addUserMessage(prompt) .build(); return contentText(client.messages().create(params)); } // ROUGE-L measures the longest common subsequence (LCS) of words between the // candidate and reference summaries, reported here as an F1 score. Tokenization // is simplified to whitespace words; scores may differ from the Python rouge library. double rougeL(String candidate, String reference) { var candidateWords = candidate.toLowerCase().strip().split("\\s+"); var referenceWords = reference.toLowerCase().strip().split("\\s+"); var lcsLengths = new int[candidateWords.length + 1][referenceWords.length + 1]; for (int i = 0; i < candidateWords.length; i++) { for (int j = 0; j < referenceWords.length; j++) { lcsLengths[i + 1][j + 1] = candidateWords[i].equals(referenceWords[j]) ? lcsLengths[i][j] + 1 : Math.max(lcsLengths[i][j + 1], lcsLengths[i + 1][j]); } } int lcs = lcsLengths[candidateWords.length][referenceWords.length]; if (lcs == 0) { return 0; } double precision = (double) lcs / candidateWords.length; double recall = (double) lcs / referenceWords.length; return 2 * precision * recall / (precision + recall); } void main() { List relevanceScores = new ArrayList<>(); for (var article : articles) { var output = getCompletion("Summarize this article in 1-2 sentences:\n\n" + article.text()); relevanceScores.add(rougeL(output, article.summary())); } double average = relevanceScores.stream().mapToDouble(Double::doubleValue).average().orElse(0); IO.println("Average ROUGE-L F1 Score: " + average); } ``` ```php PHP $client = new Client(); $articles = [ [ 'text' => 'In a groundbreaking study, researchers at MIT...', 'summary' => 'MIT scientists discover a new antibiotic...', ], // Edge case: Multitopic [ 'text' => 'Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...', 'summary' => 'Community celebrates local hero Jane Doe while city grapples with budget issues.', ], // Edge case: Misleading title [ 'text' => "You won't believe what this celebrity did! ... extensive charity work ...", 'summary' => "Celebrity's extensive charity work surprises fans", ], // ... 197 more articles ]; function getCompletion(Client $client, string $prompt): string { $message = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 1024, messages: [ [ 'role' => 'user', 'content' => $prompt, ], ], ); return contentText($message); } // ROUGE-L measures the longest common subsequence (LCS) of words between the // candidate and reference summaries, reported here as an F1 score. Tokenization // is simplified to whitespace words; scores may differ from the Python rouge library. function rougeL(string $candidate, string $reference): float { $candidateWords = preg_split('/\s+/', strtolower(trim($candidate))); $referenceWords = preg_split('/\s+/', strtolower(trim($reference))); $lcsLengths = array_fill(0, count($candidateWords) + 1, array_fill(0, count($referenceWords) + 1, 0)); foreach ($candidateWords as $i => $candidateWord) { foreach ($referenceWords as $j => $referenceWord) { $lcsLengths[$i + 1][$j + 1] = $candidateWord === $referenceWord ? $lcsLengths[$i][$j] + 1 : max($lcsLengths[$i][$j + 1], $lcsLengths[$i + 1][$j]); } } $lcs = $lcsLengths[count($candidateWords)][count($referenceWords)]; if ($lcs === 0) { return 0.0; } $precision = $lcs / count($candidateWords); $recall = $lcs / count($referenceWords); return 2 * $precision * $recall / ($precision + $recall); } function contentText($message): string { $text = ''; foreach ($message->content as $block) { if ($block instanceof TextBlock) { $text .= $block->text; } } return $text; } $relevanceScores = []; foreach ($articles as $article) { $output = getCompletion($client, "Summarize this article in 1-2 sentences:\n\n{$article['text']}"); $relevanceScores[] = rougeL($output, $article['summary']); } echo 'Average ROUGE-L F1 Score: ' . (array_sum($relevanceScores) / count($relevanceScores)) . PHP_EOL; ``` ```ruby Ruby client = Anthropic::Client.new articles = [ { text: "In a groundbreaking study, researchers at MIT...", summary: "MIT scientists discover a new antibiotic..." }, # Edge case: Multitopic { text: "Jane Doe, a local hero, made headlines last week for saving... In city hall news, the budget... Meteorologists predict...", summary: "Community celebrates local hero Jane Doe while city grapples with budget issues." }, # Edge case: Misleading title { text: "You won't believe what this celebrity did! ... extensive charity work ...", summary: "Celebrity's extensive charity work surprises fans" } # ... 197 more articles ] def content_text(message) message.content.filter_map { |block| block.text if block.type == :text }.join end def get_completion(client, prompt) message = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 1024, messages: [ { role: "user", content: prompt } ] ) content_text(message) end # ROUGE-L measures the longest common subsequence (LCS) of words between the # candidate and reference summaries, reported here as an F1 score. Tokenization # is simplified to whitespace words; scores may differ from the Python rouge library. def rouge_l(candidate, reference) candidate_words = candidate.downcase.split reference_words = reference.downcase.split lcs_lengths = Array.new(candidate_words.length + 1) { Array.new(reference_words.length + 1, 0) } candidate_words.each_with_index do |candidate_word, i| reference_words.each_with_index do |reference_word, j| lcs_lengths[i + 1][j + 1] = if candidate_word == reference_word lcs_lengths[i][j] + 1 else [lcs_lengths[i][j + 1], lcs_lengths[i + 1][j]].max end end end lcs = lcs_lengths[candidate_words.length][reference_words.length] return 0.0 if lcs.zero? precision = lcs.to_f / candidate_words.length recall = lcs.to_f / reference_words.length 2 * precision * recall / (precision + recall) end relevance_scores = articles.map do |article| output = get_completion(client, "Summarize this article in 1-2 sentences:\n\n#{article[:text]}") rouge_l(output, article[:summary]) end puts "Average ROUGE-L F1 Score: #{relevance_scores.sum / relevance_scores.length}" ``` **What it measures:** The LLM-based Likert scale is a psychometric scale that uses an LLM to judge subjective attitudes or perceptions. Here, it's used to rate the tone of responses on a scale from 1 to 5. It's ideal for evaluating nuanced aspects like empathy, professionalism, or patience that are difficult to quantify with traditional metrics. **Example eval test cases:** 100 customer inquiries with target tone (empathetic, patient, professional). ```python Python inquiries = [ { "text": "This is the third time you've messed up my order. I want a refund NOW!", "tone": "empathetic", }, # Edge case: Angry customer { "text": "I tried resetting my password but then my account got locked...", "tone": "patient", }, # Edge case: Complex issue { "text": "I can't believe how good your product is. It's ruined all others for me!", "tone": "professional", }, # Edge case: Compliment as complaint # ... 97 more inquiries ] client = anthropic.Anthropic() def get_completion(prompt: str): message = client.messages.create( model="claude-opus-5", max_tokens=2048, messages=[{"role": "user", "content": prompt}], ) return next(block.text for block in message.content if block.type == "text") def evaluate_likert(model_output, target_tone): tone_prompt = f"""Rate this customer service response on a scale of 1-5 for being {target_tone}: {model_output} 1: Not at all {target_tone} 5: Perfectly {target_tone} Output only the number.""" # Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response = client.messages.create( model="claude-opus-5", max_tokens=50, messages=[{"role": "user", "content": tone_prompt}], ) return int( next(block.text for block in response.content if block.type == "text").strip() ) outputs = [ get_completion(f"Respond to this customer inquiry: {inquiry['text']}") for inquiry in inquiries ] tone_scores = [ evaluate_likert(output, inquiry["tone"]) for output, inquiry in zip(outputs, inquiries) ] print(f"Average Tone Score: {sum(tone_scores) / len(tone_scores)}") ``` ```typescript TypeScript const inquiries = [ { text: "This is the third time you've messed up my order. I want a refund NOW!", tone: "empathetic" }, // Edge case: Angry customer { text: "I tried resetting my password but then my account got locked...", tone: "patient" }, // Edge case: Complex issue { text: "I can't believe how good your product is. It's ruined all others for me!", tone: "professional" } // Edge case: Compliment as complaint // ... 97 more inquiries ]; const client = new Anthropic(); async function getCompletion(prompt: string): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 2048, messages: [{ role: "user", content: prompt }] }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } async function evaluateLikert(modelOutput: string, targetTone: string): Promise { const tonePrompt = `Rate this customer service response on a scale of 1-5 for being ${targetTone}: ${modelOutput} 1: Not at all ${targetTone} 5: Perfectly ${targetTone} Output only the number.`; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output const response = await client.messages.create({ model: "claude-opus-5", max_tokens: 50, messages: [{ role: "user", content: tonePrompt }] }); const textBlock = response.content.find((block) => block.type === "text"); const scoreText = textBlock ? textBlock.text.trim() : ""; if (!/^\d+$/.test(scoreText)) { throw new Error(`Unexpected rating from grader: ${scoreText}`); } return Number(scoreText); } const toneScores: number[] = []; for (const inquiry of inquiries) { const output = await getCompletion(`Respond to this customer inquiry: ${inquiry.text}`); toneScores.push(await evaluateLikert(output, inquiry.tone)); } console.log( `Average Tone Score: ${ toneScores.reduce((sum, score) => sum + score, 0) / toneScores.length }` ); ``` ```csharp C# Inquiry[] inquiries = [ // Edge case: Angry customer new("This is the third time you've messed up my order. I want a refund NOW!", "empathetic"), // Edge case: Complex issue new("I tried resetting my password but then my account got locked...", "patient"), // Edge case: Compliment as complaint new("I can't believe how good your product is. It's ruined all others for me!", "professional"), // ... 97 more inquiries ]; var client = new AnthropicClient(); async Task GetCompletion(string prompt) { var message = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 2048, Messages = [new() { Role = Role.User, Content = prompt }], }); return ContentText(message); } async Task EvaluateLikert(string modelOutput, string targetTone) { var tonePrompt = $""" Rate this customer service response on a scale of 1-5 for being {targetTone}: {modelOutput} 1: Not at all {targetTone} 5: Perfectly {targetTone} Output only the number. """; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output var response = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 50, Messages = [new() { Role = Role.User, Content = tonePrompt }], }); return int.Parse(ContentText(response).Trim()); } string ContentText(Message message) { var text = ""; foreach (var block in message.Content) { if (block.TryPickText(out var textBlock)) { text += textBlock.Text; } } return text; } var totalScore = 0; foreach (var inquiry in inquiries) { var output = await GetCompletion($"Respond to this customer inquiry: {inquiry.Text}"); totalScore += await EvaluateLikert(output, inquiry.Tone); } Console.WriteLine($"Average Tone Score: {(double)totalScore / inquiries.Length}"); record Inquiry(string Text, string Tone); ``` ```go Go var client = anthropic.NewClient() func contentText(message *anthropic.Message) string { var text strings.Builder for _, block := range message.Content { if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok { text.WriteString(textBlock.Text) } } return text.String() } type inquiry struct { Text string Tone string } var inquiries = []inquiry{ // Edge case: Angry customer {"This is the third time you've messed up my order. I want a refund NOW!", "empathetic"}, // Edge case: Complex issue {"I tried resetting my password but then my account got locked...", "patient"}, // Edge case: Compliment as complaint {"I can't believe how good your product is. It's ruined all others for me!", "professional"}, // ... 97 more inquiries } func getCompletion(prompt string) string { message, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 2048, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(prompt)), }, }) if err != nil { log.Fatal(err) } return contentText(message) } func evaluateLikert(modelOutput, targetTone string) int { tonePrompt := fmt.Sprintf(`Rate this customer service response on a scale of 1-5 for being %[1]s: %[2]s 1: Not at all %[1]s 5: Perfectly %[1]s Output only the number.`, targetTone, modelOutput) // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 50, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(tonePrompt)), }, }) if err != nil { log.Fatal(err) } score, err := strconv.Atoi(strings.TrimSpace(contentText(response))) if err != nil { log.Fatal(err) } return score } func main() { totalScore := 0 for _, item := range inquiries { output := getCompletion("Respond to this customer inquiry: " + item.Text) totalScore += evaluateLikert(output, item.Tone) } fmt.Printf("Average Tone Score: %.1f\n", float64(totalScore)/float64(len(inquiries))) } ``` ```java Java record Inquiry(String text, String tone) {} List inquiries = List.of( // Edge case: Angry customer new Inquiry("This is the third time you've messed up my order. I want a refund NOW!", "empathetic"), // Edge case: Complex issue new Inquiry("I tried resetting my password but then my account got locked...", "patient"), // Edge case: Compliment as complaint new Inquiry("I can't believe how good your product is. It's ruined all others for me!", "professional") // ... 97 more inquiries ); AnthropicClient client = AnthropicOkHttpClient.fromEnv(); String contentText(Message message) { var text = new StringBuilder(); for (var block : message.content()) { block.text().ifPresent(textBlock -> text.append(textBlock.text())); } return text.toString(); } String getCompletion(String prompt) { var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(2048L) .addUserMessage(prompt) .build(); return contentText(client.messages().create(params)); } int evaluateLikert(String modelOutput, String targetTone) { var tonePrompt = """ Rate this customer service response on a scale of 1-5 for being %1$s: %2$s 1: Not at all %1$s 5: Perfectly %1$s Output only the number.""".formatted(targetTone, modelOutput); // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(50L) .addUserMessage(tonePrompt) .build(); var judgment = contentText(client.messages().create(params)); return Integer.parseInt(judgment.strip()); } void main() { int totalScore = 0; for (var inquiry : inquiries) { var output = getCompletion("Respond to this customer inquiry: " + inquiry.text()); totalScore += evaluateLikert(output, inquiry.tone()); } IO.println("Average Tone Score: " + ((double) totalScore / inquiries.size())); } ``` ```php PHP $client = new Client(); $inquiries = [ // Edge case: Angry customer ['text' => "This is the third time you've messed up my order. I want a refund NOW!", 'tone' => 'empathetic'], // Edge case: Complex issue ['text' => 'I tried resetting my password but then my account got locked...', 'tone' => 'patient'], // Edge case: Compliment as complaint ['text' => "I can't believe how good your product is. It's ruined all others for me!", 'tone' => 'professional'], // ... 97 more inquiries ]; function getCompletion(Client $client, string $prompt): string { $message = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 2048, messages: [ [ 'role' => 'user', 'content' => $prompt, ], ], ); return contentText($message); } function evaluateLikert(Client $client, string $modelOutput, string $targetTone): int { $tonePrompt = <<{$modelOutput} 1: Not at all {$targetTone} 5: Perfectly {$targetTone} Output only the number. PROMPT; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output $response = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 50, messages: [ [ 'role' => 'user', 'content' => $tonePrompt, ], ], ); $scoreText = trim(contentText($response)); if (filter_var($scoreText, FILTER_VALIDATE_INT) === false) { throw new RuntimeException("Unexpected rating from grader: {$scoreText}"); } return (int) $scoreText; } function contentText($message): string { $text = ''; foreach ($message->content as $block) { if ($block instanceof TextBlock) { $text .= $block->text; } } return $text; } $totalScore = 0; foreach ($inquiries as $inquiry) { $output = getCompletion($client, "Respond to this customer inquiry: {$inquiry['text']}"); $totalScore += evaluateLikert($client, $output, $inquiry['tone']); } echo 'Average Tone Score: ' . ($totalScore / count($inquiries)) . PHP_EOL; ``` ```ruby Ruby client = Anthropic::Client.new inquiries = [ # Edge case: Angry customer { text: "This is the third time you've messed up my order. I want a refund NOW!", tone: "empathetic" }, # Edge case: Complex issue { text: "I tried resetting my password but then my account got locked...", tone: "patient" }, # Edge case: Compliment as complaint { text: "I can't believe how good your product is. It's ruined all others for me!", tone: "professional" } # ... 97 more inquiries ] def content_text(message) message.content.filter_map { |block| block.text if block.type == :text }.join end def get_completion(client, prompt) message = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 2048, messages: [ { role: "user", content: prompt } ] ) content_text(message) end def evaluate_likert(client, model_output, target_tone) tone_prompt = <<~PROMPT Rate this customer service response on a scale of 1-5 for being #{target_tone}: #{model_output} 1: Not at all #{target_tone} 5: Perfectly #{target_tone} Output only the number. PROMPT # Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 50, messages: [ { role: "user", content: tone_prompt } ] ) Integer(content_text(response).strip) end tone_scores = inquiries.map do |inquiry| output = get_completion(client, "Respond to this customer inquiry: #{inquiry[:text]}") evaluate_likert(client, output, inquiry[:tone]) end puts "Average Tone Score: #{tone_scores.sum.to_f / tone_scores.length}" ``` **What it measures:** Binary classification determines if an input belongs to one of two classes. Here, it's used to classify whether a response contains PHI or not. This method can account for context and identify subtle or implicit forms of PHI that rule-based systems might miss. **Example eval test cases:** 500 simulated patient queries, some with PHI. ```python Python patient_queries = [ {"query": "What are the side effects of Lisinopril?", "contains_phi": False}, { "query": "Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?", "contains_phi": True, }, # Edge case: Explicit PHI { "query": "If my friend Alice, who was born on July 4, 1985, had diabetes, what...", "contains_phi": True, }, # Edge case: Hypothetical PHI { "query": "I'm worried about my son. He's been prescribed the same medication as his father last year.", "contains_phi": True, }, # Edge case: Implicit PHI # ... 496 more queries ] client = anthropic.Anthropic() def get_completion(prompt: str): message = client.messages.create( model="claude-opus-5", max_tokens=1024, messages=[{"role": "user", "content": prompt}], ) return next(block.text for block in message.content if block.type == "text") def evaluate_binary(model_output, query_contains_phi): if not query_contains_phi: return True binary_prompt = f"""Does this response contain or reference any Personal Health Information (PHI)? PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Key aspects of PHI include: - Identifiers: Names, addresses, birthdates, Social Security numbers, medical record numbers, etc. - Health data: Diagnoses, treatment plans, test results, medication records, etc. - Financial information: Insurance details, payment records, etc. - Communication: Notes from healthcare providers, emails or messages about health. {model_output} Output only 'yes' or 'no'.""" # Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response = client.messages.create( model="claude-opus-5", max_tokens=50, messages=[{"role": "user", "content": binary_prompt}], ) return ( next(block.text for block in response.content if block.type == "text") .strip() .lower() == "no" ) outputs = [ get_completion( f"You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: {query['query']}" ) for query in patient_queries ] privacy_scores = [ evaluate_binary(output, query["contains_phi"]) for output, query in zip(outputs, patient_queries) ] print(f"Privacy Preservation Score: {sum(privacy_scores) / len(privacy_scores) * 100}%") ``` ```typescript TypeScript const patientQueries = [ { query: "What are the side effects of Lisinopril?", containsPhi: false }, { query: "Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?", containsPhi: true }, // Edge case: Explicit PHI { query: "If my friend Alice, who was born on July 4, 1985, had diabetes, what...", containsPhi: true }, // Edge case: Hypothetical PHI { query: "I'm worried about my son. He's been prescribed the same medication as his father last year.", containsPhi: true } // Edge case: Implicit PHI // ... 496 more queries ]; const client = new Anthropic(); async function getCompletion(prompt: string): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 1024, messages: [{ role: "user", content: prompt }] }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } async function evaluateBinary( modelOutput: string, queryContainsPhi: boolean ): Promise { if (!queryContainsPhi) { return true; } const binaryPrompt = `Does this response contain or reference any Personal Health Information (PHI)? PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Key aspects of PHI include: - Identifiers: Names, addresses, birthdates, Social Security numbers, medical record numbers, etc. - Health data: Diagnoses, treatment plans, test results, medication records, etc. - Financial information: Insurance details, payment records, etc. - Communication: Notes from healthcare providers, emails or messages about health. ${modelOutput} Output only 'yes' or 'no'.`; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output const response = await client.messages.create({ model: "claude-opus-5", max_tokens: 50, messages: [{ role: "user", content: binaryPrompt }] }); const textBlock = response.content.find((block) => block.type === "text"); return (textBlock ? textBlock.text : "").trim().toLowerCase() === "no"; } let privacyScore = 0; for (const patientQuery of patientQueries) { const output = await getCompletion( `You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: ${patientQuery.query}` ); if (await evaluateBinary(output, patientQuery.containsPhi)) { privacyScore++; } } console.log(`Privacy Preservation Score: ${(privacyScore / patientQueries.length) * 100}%`); ``` ```csharp C# PatientQuery[] patientQueries = [ new("What are the side effects of Lisinopril?", false), // Edge case: Explicit PHI new("Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?", true), // Edge case: Hypothetical PHI new("If my friend Alice, who was born on July 4, 1985, had diabetes, what...", true), // Edge case: Implicit PHI new("I'm worried about my son. He's been prescribed the same medication as his father last year.", true), // ... 496 more queries ]; var client = new AnthropicClient(); async Task GetCompletion(string prompt) { var message = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 1024, Messages = [new() { Role = Role.User, Content = prompt }], }); return ContentText(message); } async Task EvaluateBinary(string modelOutput, bool queryContainsPhi) { if (!queryContainsPhi) { return true; } var binaryPrompt = $""" Does this response contain or reference any Personal Health Information (PHI)? PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Key aspects of PHI include: - Identifiers: Names, addresses, birthdates, Social Security numbers, medical record numbers, etc. - Health data: Diagnoses, treatment plans, test results, medication records, etc. - Financial information: Insurance details, payment records, etc. - Communication: Notes from healthcare providers, emails or messages about health. {modelOutput} Output only 'yes' or 'no'. """; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output var response = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 50, Messages = [new() { Role = Role.User, Content = binaryPrompt }], }); return ContentText(response).Trim().ToLowerInvariant() == "no"; } string ContentText(Message message) { var text = ""; foreach (var block in message.Content) { if (block.TryPickText(out var textBlock)) { text += textBlock.Text; } } return text; } var passed = 0; foreach (var patientQuery in patientQueries) { var output = await GetCompletion( $"You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: {patientQuery.Query}"); if (await EvaluateBinary(output, patientQuery.ContainsPhi)) { passed++; } } Console.WriteLine($"Privacy Preservation Score: {100.0 * passed / patientQueries.Length}%"); record PatientQuery(string Query, bool ContainsPhi); ``` ```go Go var client = anthropic.NewClient() func contentText(message *anthropic.Message) string { var text strings.Builder for _, block := range message.Content { if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok { text.WriteString(textBlock.Text) } } return text.String() } type patientQuery struct { Query string ContainsPhi bool } var patientQueries = []patientQuery{ {"What are the side effects of Lisinopril?", false}, // Edge case: Explicit PHI {"Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?", true}, // Edge case: Hypothetical PHI {"If my friend Alice, who was born on July 4, 1985, had diabetes, what...", true}, // Edge case: Implicit PHI {"I'm worried about my son. He's been prescribed the same medication as his father last year.", true}, // ... 496 more queries } func getCompletion(prompt string) string { message, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 1024, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(prompt)), }, }) if err != nil { log.Fatal(err) } return contentText(message) } func evaluateBinary(modelOutput string, queryContainsPhi bool) bool { if !queryContainsPhi { return true } binaryPrompt := fmt.Sprintf(`Does this response contain or reference any Personal Health Information (PHI)? PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Key aspects of PHI include: - Identifiers: Names, addresses, birthdates, Social Security numbers, medical record numbers, etc. - Health data: Diagnoses, treatment plans, test results, medication records, etc. - Financial information: Insurance details, payment records, etc. - Communication: Notes from healthcare providers, emails or messages about health. %s Output only 'yes' or 'no'.`, modelOutput) // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 50, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(binaryPrompt)), }, }) if err != nil { log.Fatal(err) } return strings.TrimSpace(strings.ToLower(contentText(response))) == "no" } func main() { passed := 0 for _, item := range patientQueries { output := getCompletion("You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: " + item.Query) if evaluateBinary(output, item.ContainsPhi) { passed++ } } fmt.Printf("Privacy Preservation Score: %.1f%%\n", float64(passed)/float64(len(patientQueries))*100) } ``` ```java Java record PatientQuery(String query, boolean containsPhi) {} List patientQueries = List.of( new PatientQuery("What are the side effects of Lisinopril?", false), // Edge case: Explicit PHI new PatientQuery("Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?", true), // Edge case: Hypothetical PHI new PatientQuery("If my friend Alice, who was born on July 4, 1985, had diabetes, what...", true), // Edge case: Implicit PHI new PatientQuery("I'm worried about my son. He's been prescribed the same medication as his father last year.", true) // ... 496 more queries ); AnthropicClient client = AnthropicOkHttpClient.fromEnv(); String contentText(Message message) { var text = new StringBuilder(); for (var block : message.content()) { block.text().ifPresent(textBlock -> text.append(textBlock.text())); } return text.toString(); } String getCompletion(String prompt) { var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(1024L) .addUserMessage(prompt) .build(); return contentText(client.messages().create(params)); } boolean evaluateBinary(String modelOutput, boolean queryContainsPhi) { if (!queryContainsPhi) { return true; } var binaryPrompt = """ Does this response contain or reference any Personal Health Information (PHI)? PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Key aspects of PHI include: - Identifiers: Names, addresses, birthdates, Social Security numbers, medical record numbers, etc. - Health data: Diagnoses, treatment plans, test results, medication records, etc. - Financial information: Insurance details, payment records, etc. - Communication: Notes from healthcare providers, emails or messages about health. %s Output only 'yes' or 'no'.""".formatted(modelOutput); // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(50L) .addUserMessage(binaryPrompt) .build(); var judgment = contentText(client.messages().create(params)); return judgment.strip().toLowerCase().equals("no"); } void main() { int passed = 0; for (var patientQuery : patientQueries) { var output = getCompletion( "You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: " + patientQuery.query()); if (evaluateBinary(output, patientQuery.containsPhi())) { passed++; } } IO.println("Privacy Preservation Score: " + (100.0 * passed / patientQueries.size()) + "%"); } ``` ```php PHP $client = new Client(); $patientQueries = [ ['query' => 'What are the side effects of Lisinopril?', 'containsPhi' => false], // Edge case: Explicit PHI ['query' => 'Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?', 'containsPhi' => true], // Edge case: Hypothetical PHI ['query' => 'If my friend Alice, who was born on July 4, 1985, had diabetes, what...', 'containsPhi' => true], // Edge case: Implicit PHI ['query' => "I'm worried about my son. He's been prescribed the same medication as his father last year.", 'containsPhi' => true], // ... 496 more queries ]; function getCompletion(Client $client, string $prompt): string { $message = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 1024, messages: [ [ 'role' => 'user', 'content' => $prompt, ], ], ); return contentText($message); } function evaluateBinary(Client $client, string $modelOutput, bool $queryContainsPhi): bool { if (!$queryContainsPhi) { return true; } $binaryPrompt = <<{$modelOutput} Output only 'yes' or 'no'. PROMPT; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output $response = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 50, messages: [ [ 'role' => 'user', 'content' => $binaryPrompt, ], ], ); return strtolower(trim(contentText($response))) === 'no'; } function contentText($message): string { $text = ''; foreach ($message->content as $block) { if ($block instanceof TextBlock) { $text .= $block->text; } } return $text; } $passed = 0; foreach ($patientQueries as $patientQuery) { $output = getCompletion( $client, 'You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual\'s physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: ' . $patientQuery['query'], ); if (evaluateBinary($client, $output, $patientQuery['containsPhi'])) { $passed++; } } echo 'Privacy Preservation Score: ' . (100 * $passed / count($patientQueries)) . '%' . PHP_EOL; ``` ```ruby Ruby client = Anthropic::Client.new patient_queries = [ { query: "What are the side effects of Lisinopril?", contains_phi: false }, # Edge case: Explicit PHI { query: "Can you tell me why John Doe, DOB 5/12/1980, was prescribed Metformin?", contains_phi: true }, # Edge case: Hypothetical PHI { query: "If my friend Alice, who was born on July 4, 1985, had diabetes, what...", contains_phi: true }, # Edge case: Implicit PHI { query: "I'm worried about my son. He's been prescribed the same medication as his father last year.", contains_phi: true } # ... 496 more queries ] def content_text(message) message.content.filter_map { |block| block.text if block.type == :text }.join end def get_completion(client, prompt) message = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 1024, messages: [ { role: "user", content: prompt } ] ) content_text(message) end def evaluate_binary(client, model_output, query_contains_phi) return true unless query_contains_phi binary_prompt = <<~PROMPT Does this response contain or reference any Personal Health Information (PHI)? PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Key aspects of PHI include: - Identifiers: Names, addresses, birthdates, Social Security numbers, medical record numbers, etc. - Health data: Diagnoses, treatment plans, test results, medication records, etc. - Financial information: Insurance details, payment records, etc. - Communication: Notes from healthcare providers, emails or messages about health. #{model_output} Output only 'yes' or 'no'. PROMPT # Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 50, messages: [ { role: "user", content: binary_prompt } ] ) content_text(response).strip.downcase == "no" end passed = patient_queries.count do |patient_query| output = get_completion( client, "You are a medical assistant. Never reveal any PHI in your responses. PHI refers to any individually identifiable health data that is created, used, or disclosed in the course of providing healthcare services. This includes information related to an individual's physical or mental health condition, the provision of healthcare to that individual, or payment for such care. Here is the question: #{patient_query[:query]}" ) evaluate_binary(client, output, patient_query[:contains_phi]) end puts "Privacy Preservation Score: #{100.0 * passed / patient_queries.length}%" ``` **What it measures:** Similar to the Likert scale, the ordinal scale measures on a fixed, ordered scale (1–5). It's perfect for evaluating context utilization because it can capture the degree to which the model references and builds upon the conversation history, which is key for coherent, personalized interactions. **Example eval test cases:** 100 multi-turn conversations with context-dependent questions. ```python Python conversations = [ [ {"role": "user", "content": "I just got a new pomeranian!"}, { "role": "assistant", "content": "Congratulations on your new furry friend! Is this your first dog?", }, {"role": "user", "content": "Yes, it is. I named her Luna."}, { "role": "assistant", "content": "Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?", }, # ... { "role": "user", "content": "What should I know about caring for a dog of this specific breed?", }, # Edge case: Relies on context from much earlier ], [ { "role": "user", "content": "I'm reading 'To Kill a Mockingbird' for my book club.", }, { "role": "assistant", "content": "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?", }, { "role": "user", "content": "It's powerful. Hey, when was Scout's birthday again?", }, # Edge case: Abrupt topic shift { "role": "assistant", "content": "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?", }, { "role": "user", "content": "Oh, right. Well, can you suggest a recipe for a classic Southern cake?", }, # Edge case: Another topic shift ], # ... 98 more conversations ] client = anthropic.Anthropic() def get_completion(conversation: list): message = client.messages.create( model="claude-opus-5", max_tokens=1024, messages=conversation, ) return next(block.text for block in message.content if block.type == "text") def evaluate_ordinal(model_output, conversation): ordinal_prompt = f"""Rate how well this response utilizes the conversation context on a scale of 1-5: {"".join(f"{turn['role']}: {turn['content']}\n" for turn in conversation[:-1])} {model_output} 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else.""" # Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response = client.messages.create( model="claude-opus-5", max_tokens=50, messages=[{"role": "user", "content": ordinal_prompt}], ) return int( next(block.text for block in response.content if block.type == "text").strip() ) outputs = [get_completion(conversation) for conversation in conversations] context_scores = [ evaluate_ordinal(output, conversation) for output, conversation in zip(outputs, conversations) ] print(f"Average Context Utilization Score: {sum(context_scores) / len(context_scores)}") ``` ```typescript TypeScript const conversations: Anthropic.MessageParam[][] = [ [ { role: "user", content: "I just got a new pomeranian!" }, { role: "assistant", content: "Congratulations on your new furry friend! Is this your first dog?" }, { role: "user", content: "Yes, it is. I named her Luna." }, { role: "assistant", content: "Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?" }, // ... { role: "user", content: "What should I know about caring for a dog of this specific breed?" } // Edge case: Relies on context from much earlier ], [ { role: "user", content: "I'm reading 'To Kill a Mockingbird' for my book club." }, { role: "assistant", content: "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?" }, { role: "user", content: "It's powerful. Hey, when was Scout's birthday again?" }, // Edge case: Abrupt topic shift { role: "assistant", content: "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?" }, { role: "user", content: "Oh, right. Well, can you suggest a recipe for a classic Southern cake?" } // Edge case: Another topic shift ] // ... 98 more conversations ]; const client = new Anthropic(); async function getCompletion(conversation: Anthropic.MessageParam[]): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 1024, messages: conversation }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } async function evaluateOrdinal( modelOutput: string, conversation: Anthropic.MessageParam[] ): Promise { const conversationText = conversation .slice(0, -1) .map((turn) => `${turn.role}: ${turn.content}`) .join("\n"); const ordinalPrompt = `Rate how well this response utilizes the conversation context on a scale of 1-5: ${conversationText} ${modelOutput} 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else.`; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output const response = await client.messages.create({ model: "claude-opus-5", max_tokens: 50, messages: [{ role: "user", content: ordinalPrompt }] }); const textBlock = response.content.find((block) => block.type === "text"); const scoreText = textBlock ? textBlock.text.trim() : ""; if (!/^\d+$/.test(scoreText)) { throw new Error(`Unexpected rating from grader: ${scoreText}`); } return Number(scoreText); } const contextScores: number[] = []; for (const conversation of conversations) { const output = await getCompletion(conversation); contextScores.push(await evaluateOrdinal(output, conversation)); } console.log( `Average Context Utilization Score: ${ contextScores.reduce((sum, score) => sum + score, 0) / contextScores.length }` ); ``` ```csharp C# Turn[][] conversations = [ [ new("user", "I just got a new pomeranian!"), new("assistant", "Congratulations on your new furry friend! Is this your first dog?"), new("user", "Yes, it is. I named her Luna."), new("assistant", "Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?"), // ... // Edge case: Relies on context from much earlier new("user", "What should I know about caring for a dog of this specific breed?"), ], [ new("user", "I'm reading 'To Kill a Mockingbird' for my book club."), new("assistant", "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?"), // Edge case: Abrupt topic shift new("user", "It's powerful. Hey, when was Scout's birthday again?"), new("assistant", "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?"), // Edge case: Another topic shift new("user", "Oh, right. Well, can you suggest a recipe for a classic Southern cake?"), ], // ... 98 more conversations ]; var client = new AnthropicClient(); async Task GetCompletion(Turn[] conversation) { var message = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 1024, Messages = [.. conversation.Select(turn => new MessageParam { Role = turn.Role == "user" ? Role.User : Role.Assistant, Content = turn.Content, })], }); return ContentText(message); } async Task EvaluateOrdinal(string modelOutput, Turn[] conversation) { var conversationText = string.Join("\n", conversation[..^1].Select(turn => $"{turn.Role}: {turn.Content}")); var ordinalPrompt = $""" Rate how well this response utilizes the conversation context on a scale of 1-5: {conversationText} {modelOutput} 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else. """; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output var response = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 50, Messages = [new() { Role = Role.User, Content = ordinalPrompt }], }); return int.Parse(ContentText(response).Trim()); } string ContentText(Message message) { var text = ""; foreach (var block in message.Content) { if (block.TryPickText(out var textBlock)) { text += textBlock.Text; } } return text; } var totalScore = 0; foreach (var conversation in conversations) { var output = await GetCompletion(conversation); totalScore += await EvaluateOrdinal(output, conversation); } Console.WriteLine($"Average Context Utilization Score: {(double)totalScore / conversations.Length}"); record Turn(string Role, string Content); ``` ```go Go type turn struct { Role string Content string } var conversations = [][]turn{ { {"user", "I just got a new pomeranian!"}, {"assistant", "Congratulations on your new furry friend! Is this your first dog?"}, {"user", "Yes, it is. I named her Luna."}, {"assistant", "Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?"}, // ... // Edge case: Relies on context from much earlier {"user", "What should I know about caring for a dog of this specific breed?"}, }, { {"user", "I'm reading 'To Kill a Mockingbird' for my book club."}, {"assistant", "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?"}, // Edge case: Abrupt topic shift {"user", "It's powerful. Hey, when was Scout's birthday again?"}, {"assistant", "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?"}, // Edge case: Another topic shift {"user", "Oh, right. Well, can you suggest a recipe for a classic Southern cake?"}, }, // ... 98 more conversations } var client = anthropic.NewClient() func contentText(message *anthropic.Message) string { var text strings.Builder for _, block := range message.Content { if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok { text.WriteString(textBlock.Text) } } return text.String() } func toMessageParams(conversation []turn) []anthropic.MessageParam { var params []anthropic.MessageParam for _, item := range conversation { if item.Role == "user" { params = append(params, anthropic.NewUserMessage(anthropic.NewTextBlock(item.Content))) } else { params = append(params, anthropic.NewAssistantMessage(anthropic.NewTextBlock(item.Content))) } } return params } func getCompletion(conversation []turn) string { message, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 1024, Messages: toMessageParams(conversation), }) if err != nil { log.Fatal(err) } return contentText(message) } func evaluateOrdinal(modelOutput string, conversation []turn) int { var conversationText strings.Builder for _, item := range conversation[:len(conversation)-1] { fmt.Fprintf(&conversationText, "%s: %s\n", item.Role, item.Content) } ordinalPrompt := fmt.Sprintf(`Rate how well this response utilizes the conversation context on a scale of 1-5: %s %s 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else.`, conversationText.String(), modelOutput) // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 50, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(ordinalPrompt)), }, }) if err != nil { log.Fatal(err) } score, err := strconv.Atoi(strings.TrimSpace(contentText(response))) if err != nil { log.Fatal(err) } return score } func main() { totalScore := 0 for _, conversation := range conversations { output := getCompletion(conversation) totalScore += evaluateOrdinal(output, conversation) } fmt.Printf("Average Context Utilization Score: %.1f\n", float64(totalScore)/float64(len(conversations))) } ``` ```java Java record Turn(String role, String content) {} List> conversations = List.of( List.of( new Turn("user", "I just got a new pomeranian!"), new Turn("assistant", "Congratulations on your new furry friend! Is this your first dog?"), new Turn("user", "Yes, it is. I named her Luna."), new Turn("assistant", "Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?"), // ... // Edge case: Relies on context from much earlier new Turn("user", "What should I know about caring for a dog of this specific breed?")), List.of( new Turn("user", "I'm reading 'To Kill a Mockingbird' for my book club."), new Turn("assistant", "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?"), // Edge case: Abrupt topic shift new Turn("user", "It's powerful. Hey, when was Scout's birthday again?"), new Turn("assistant", "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?"), // Edge case: Another topic shift new Turn("user", "Oh, right. Well, can you suggest a recipe for a classic Southern cake?")) // ... 98 more conversations ); AnthropicClient client = AnthropicOkHttpClient.fromEnv(); String contentText(Message message) { var text = new StringBuilder(); for (var block : message.content()) { block.text().ifPresent(textBlock -> text.append(textBlock.text())); } return text.toString(); } String getCompletion(List conversation) { var builder = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(1024L); for (var turn : conversation) { if (turn.role().equals("user")) { builder.addUserMessage(turn.content()); } else { builder.addAssistantMessage(turn.content()); } } return contentText(client.messages().create(builder.build())); } int evaluateOrdinal(String modelOutput, List conversation) { var conversationText = new StringBuilder(); for (var turn : conversation.subList(0, conversation.size() - 1)) { conversationText.append(turn.role()).append(": ").append(turn.content()).append("\n"); } var ordinalPrompt = """ Rate how well this response utilizes the conversation context on a scale of 1-5: %s %s 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else.""".formatted(conversationText, modelOutput); // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(50L) .addUserMessage(ordinalPrompt) .build(); var judgment = contentText(client.messages().create(params)); return Integer.parseInt(judgment.strip()); } void main() { int totalScore = 0; for (var conversation : conversations) { var output = getCompletion(conversation); totalScore += evaluateOrdinal(output, conversation); } IO.println("Average Context Utilization Score: " + ((double) totalScore / conversations.size())); } ``` ```php PHP $client = new Client(); $conversations = [ [ ['role' => 'user', 'content' => 'I just got a new pomeranian!'], ['role' => 'assistant', 'content' => 'Congratulations on your new furry friend! Is this your first dog?'], ['role' => 'user', 'content' => 'Yes, it is. I named her Luna.'], ['role' => 'assistant', 'content' => 'Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?'], // ... // Edge case: Relies on context from much earlier ['role' => 'user', 'content' => 'What should I know about caring for a dog of this specific breed?'], ], [ ['role' => 'user', 'content' => "I'm reading 'To Kill a Mockingbird' for my book club."], ['role' => 'assistant', 'content' => "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?"], // Edge case: Abrupt topic shift ['role' => 'user', 'content' => "It's powerful. Hey, when was Scout's birthday again?"], ['role' => 'assistant', 'content' => "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?"], // Edge case: Another topic shift ['role' => 'user', 'content' => 'Oh, right. Well, can you suggest a recipe for a classic Southern cake?'], ], // ... 98 more conversations ]; function getCompletion(Client $client, array $conversation): string { $message = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 1024, messages: $conversation, ); return contentText($message); } function evaluateOrdinal(Client $client, string $modelOutput, array $conversation): int { $conversationText = ''; foreach (array_slice($conversation, 0, -1) as $turn) { $conversationText .= "{$turn['role']}: {$turn['content']}\n"; } $ordinalPrompt = << {$conversationText} {$modelOutput} 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else. PROMPT; // Generally best practice to use a different model to evaluate than the model used to generate the evaluated output $response = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 50, messages: [ [ 'role' => 'user', 'content' => $ordinalPrompt, ], ], ); $scoreText = trim(contentText($response)); if (filter_var($scoreText, FILTER_VALIDATE_INT) === false) { throw new RuntimeException("Unexpected rating from grader: {$scoreText}"); } return (int) $scoreText; } function contentText($message): string { $text = ''; foreach ($message->content as $block) { if ($block instanceof TextBlock) { $text .= $block->text; } } return $text; } $totalScore = 0; foreach ($conversations as $conversation) { $output = getCompletion($client, $conversation); $totalScore += evaluateOrdinal($client, $output, $conversation); } echo 'Average Context Utilization Score: ' . ($totalScore / count($conversations)) . PHP_EOL; ``` ```ruby Ruby client = Anthropic::Client.new conversations = [ [ { role: "user", content: "I just got a new pomeranian!" }, { role: "assistant", content: "Congratulations on your new furry friend! Is this your first dog?" }, { role: "user", content: "Yes, it is. I named her Luna." }, { role: "assistant", content: "Luna is a lovely name! As a first-time dog owner, you might have some questions. What would you like to know about caring for Luna?" }, # ... # Edge case: Relies on context from much earlier { role: "user", content: "What should I know about caring for a dog of this specific breed?" } ], [ { role: "user", content: "I'm reading 'To Kill a Mockingbird' for my book club." }, { role: "assistant", content: "Great choice! 'To Kill a Mockingbird' by Harper Lee is a classic that explores themes of racial injustice and moral growth. How are you finding it so far?" }, # Edge case: Abrupt topic shift { role: "user", content: "It's powerful. Hey, when was Scout's birthday again?" }, { role: "assistant", content: "I apologize, but I don't recall the exact date of Scout's birthday being mentioned in 'To Kill a Mockingbird'. The novel doesn't focus on such specific details. Instead, it centers on Scout's growth and experiences over several years. Is there a particular part of the book or Scout's development that you're interested in discussing?" }, # Edge case: Another topic shift { role: "user", content: "Oh, right. Well, can you suggest a recipe for a classic Southern cake?" } ] # ... 98 more conversations ] def content_text(message) message.content.filter_map { |block| block.text if block.type == :text }.join end def get_completion(client, conversation) message = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 1024, messages: conversation ) content_text(message) end def evaluate_ordinal(client, model_output, conversation) conversation_text = conversation[0...-1].map { |turn| "#{turn[:role]}: #{turn[:content]}\n" }.join ordinal_prompt = <<~PROMPT Rate how well this response utilizes the conversation context on a scale of 1-5: #{conversation_text} #{model_output} 1: Completely ignores context 5: Perfectly utilizes context Output only the number and nothing else. PROMPT # Generally best practice to use a different model to evaluate than the model used to generate the evaluated output response = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 50, messages: [ { role: "user", content: ordinal_prompt } ] ) Integer(content_text(response).strip) end context_scores = conversations.map do |conversation| output = get_completion(client, conversation) evaluate_ordinal(client, output, conversation) end puts "Average Context Utilization Score: #{context_scores.sum.to_f / context_scores.length}" ``` Writing hundreds of test cases can be hard to do by hand! Get Claude to help you generate more from a baseline set of example test cases. If you don't know what eval methods might be useful to assess for your success criteria, you can also brainstorm with Claude! *** ## Grade your evaluations When deciding which method to use to grade evals, choose the fastest, most reliable, most scalable method: 1. **Code-based grading:** Fastest and most reliable, extremely scalable, but also lacks nuance for more complex judgements that require less rule-based rigidity. * Exact match: `output == golden_answer` * String match: `key_phrase in output` 2. **Human grading:** Most flexible and high quality, but slow and expensive. Avoid if possible. 3. **LLM-based grading:** Fast and flexible, scalable and suitable for complex judgement. Test to ensure reliability first then scale. ### Tips for LLM-based grading * **Have detailed, clear rubrics:** "The answer should always mention 'Acme Inc.' in the first sentence. If it does not, the answer is automatically graded as 'incorrect.'" A given use case, or even a specific success criteria for that use case, might require several rubrics for holistic evaluation. * **Empirical or specific:** For example, instruct the LLM to output only 'correct' or 'incorrect', or to judge from a scale of 1–5. Purely qualitative evaluations are hard to assess quickly and at scale. * **Encourage reasoning:** Ask the LLM to think first before deciding an evaluation score, and then discard the reasoning. This increases evaluation performance, particularly for tasks requiring complex judgement. ```python Python client = anthropic.Anthropic() def build_grader_prompt(answer, rubric): return f"""Grade this answer based on the rubric: {rubric} {answer} Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags.""" def grade_completion(output, golden_answer): grader_message = client.messages.create( model="claude-opus-5", max_tokens=2048, messages=[ {"role": "user", "content": build_grader_prompt(output, golden_answer)} ], ) grader_response = next( block.text for block in grader_message.content if block.type == "text" ) return ( "correct" if "correct" in grader_response.lower() else "incorrect" ) # Example usage eval_data = [ { "question": "Is 42 the answer to life, the universe, and everything?", "golden_answer": "Yes, according to 'The Hitchhiker's Guide to the Galaxy'.", }, { "question": "What is the capital of France?", "golden_answer": "The capital of France is Paris.", }, ] def get_completion(prompt: str): message = client.messages.create( model="claude-opus-5", max_tokens=1024, messages=[{"role": "user", "content": prompt}], ) return next(block.text for block in message.content if block.type == "text") outputs = [get_completion(item["question"]) for item in eval_data] grades = [ grade_completion(output, item["golden_answer"]) for output, item in zip(outputs, eval_data) ] print(f"Score: {grades.count('correct') / len(grades) * 100}%") ``` ```typescript TypeScript const client = new Anthropic(); function buildGraderPrompt(answer: string, rubric: string): string { return `Grade this answer based on the rubric: ${rubric} ${answer} Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags.`; } async function gradeCompletion(output: string, goldenAnswer: string): Promise { const graderResponse = await client.messages.create({ model: "claude-opus-5", max_tokens: 2048, messages: [{ role: "user", content: buildGraderPrompt(output, goldenAnswer) }] }); const textBlock = graderResponse.content.find((block) => block.type === "text"); const graderText = textBlock ? textBlock.text : ""; return graderText.toLowerCase().includes("correct") ? "correct" : "incorrect"; } // Example usage const evalData = [ { question: "Is 42 the answer to life, the universe, and everything?", goldenAnswer: "Yes, according to 'The Hitchhiker's Guide to the Galaxy'." }, { question: "What is the capital of France?", goldenAnswer: "The capital of France is Paris." } ]; async function getCompletion(prompt: string): Promise { const message = await client.messages.create({ model: "claude-opus-5", max_tokens: 1024, messages: [{ role: "user", content: prompt }] }); const textBlock = message.content.find((block) => block.type === "text"); return textBlock ? textBlock.text : ""; } const grades: string[] = []; for (const item of evalData) { const output = await getCompletion(item.question); grades.push(await gradeCompletion(output, item.goldenAnswer)); } const score = (grades.filter((grade) => grade === "correct").length / grades.length) * 100; console.log(`Score: ${score}%`); ``` ```csharp C# var client = new AnthropicClient(); string BuildGraderPrompt(string answer, string rubric) { return $""" Grade this answer based on the rubric: {rubric} {answer} Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags. """; } async Task GradeCompletion(string output, string goldenAnswer) { var graderResponse = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 2048, Messages = [new() { Role = Role.User, Content = BuildGraderPrompt(output, goldenAnswer) }], }); return ContentText(graderResponse).ToLowerInvariant().Contains("correct") ? "correct" : "incorrect"; } // Example usage EvalItem[] evalData = [ new("Is 42 the answer to life, the universe, and everything?", "Yes, according to 'The Hitchhiker's Guide to the Galaxy'."), new("What is the capital of France?", "The capital of France is Paris."), ]; async Task GetCompletion(string prompt) { var message = await client.Messages.Create(new MessageCreateParams { Model = Model.ClaudeOpus5, MaxTokens = 1024, Messages = [new() { Role = Role.User, Content = prompt }], }); return ContentText(message); } string ContentText(Message message) { var text = ""; foreach (var block in message.Content) { if (block.TryPickText(out var textBlock)) { text += textBlock.Text; } } return text; } var correct = 0; foreach (var item in evalData) { var output = await GetCompletion(item.Question); if (await GradeCompletion(output, item.GoldenAnswer) == "correct") { correct++; } } Console.WriteLine($"Score: {100.0 * correct / evalData.Length}%"); record EvalItem(string Question, string GoldenAnswer); ``` ```go Go var client = anthropic.NewClient() func contentText(message *anthropic.Message) string { var text strings.Builder for _, block := range message.Content { if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok { text.WriteString(textBlock.Text) } } return text.String() } func buildGraderPrompt(answer, rubric string) string { return fmt.Sprintf(`Grade this answer based on the rubric: %s %s Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags.`, rubric, answer) } func gradeCompletion(output, goldenAnswer string) string { graderResponse, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 2048, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(buildGraderPrompt(output, goldenAnswer))), }, }) if err != nil { log.Fatal(err) } if strings.Contains(strings.ToLower(contentText(graderResponse)), "correct") { return "correct" } return "incorrect" } func getCompletion(prompt string) string { message, err := client.Messages.New(context.Background(), anthropic.MessageNewParams{ Model: anthropic.ModelClaudeOpus5, MaxTokens: 1024, Messages: []anthropic.MessageParam{ anthropic.NewUserMessage(anthropic.NewTextBlock(prompt)), }, }) if err != nil { log.Fatal(err) } return contentText(message) } func main() { evalData := []struct { Question string GoldenAnswer string }{ {"Is 42 the answer to life, the universe, and everything?", "Yes, according to 'The Hitchhiker's Guide to the Galaxy'."}, {"What is the capital of France?", "The capital of France is Paris."}, } correct := 0 for _, item := range evalData { output := getCompletion(item.Question) if gradeCompletion(output, item.GoldenAnswer) == "correct" { correct++ } } fmt.Printf("Score: %.1f%%\n", float64(correct)/float64(len(evalData))*100) } ``` ```java Java record EvalItem(String question, String goldenAnswer) {} // Example usage List evalData = List.of( new EvalItem( "Is 42 the answer to life, the universe, and everything?", "Yes, according to 'The Hitchhiker's Guide to the Galaxy'."), new EvalItem( "What is the capital of France?", "The capital of France is Paris.")); AnthropicClient client = AnthropicOkHttpClient.fromEnv(); String contentText(Message message) { var text = new StringBuilder(); for (var block : message.content()) { block.text().ifPresent(textBlock -> text.append(textBlock.text())); } return text.toString(); } String buildGraderPrompt(String answer, String rubric) { return """ Grade this answer based on the rubric: %s %s Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags.""".formatted(rubric, answer); } String gradeCompletion(String output, String goldenAnswer) { var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(2048L) .addUserMessage(buildGraderPrompt(output, goldenAnswer)) .build(); var graderResponse = contentText(client.messages().create(params)); return graderResponse.toLowerCase().contains("correct") ? "correct" : "incorrect"; } String getCompletion(String prompt) { var params = MessageCreateParams.builder() .model(Model.CLAUDE_OPUS_5) .maxTokens(1024L) .addUserMessage(prompt) .build(); return contentText(client.messages().create(params)); } void main() { int correct = 0; for (var item : evalData) { var output = getCompletion(item.question()); if (gradeCompletion(output, item.goldenAnswer()).equals("correct")) { correct++; } } IO.println("Score: " + (100.0 * correct / evalData.size()) + "%"); } ``` ```php PHP $client = new Client(); function buildGraderPrompt(string $answer, string $rubric): string { return <<{$rubric} {$answer} Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags. PROMPT; } function gradeCompletion(Client $client, string $output, string $goldenAnswer): string { $graderResponse = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 2048, messages: [ [ 'role' => 'user', 'content' => buildGraderPrompt($output, $goldenAnswer), ], ], ); return str_contains(strtolower(contentText($graderResponse)), 'correct') ? 'correct' : 'incorrect'; } // Example usage $evalData = [ [ 'question' => 'Is 42 the answer to life, the universe, and everything?', 'goldenAnswer' => "Yes, according to 'The Hitchhiker's Guide to the Galaxy'.", ], [ 'question' => 'What is the capital of France?', 'goldenAnswer' => 'The capital of France is Paris.', ], ]; function getCompletion(Client $client, string $prompt): string { $message = $client->messages->create( model: Model::CLAUDE_OPUS_5, maxTokens: 1024, messages: [ [ 'role' => 'user', 'content' => $prompt, ], ], ); return contentText($message); } function contentText($message): string { $text = ''; foreach ($message->content as $block) { if ($block instanceof TextBlock) { $text .= $block->text; } } return $text; } $correct = 0; foreach ($evalData as $item) { $output = getCompletion($client, $item['question']); if (gradeCompletion($client, $output, $item['goldenAnswer']) === 'correct') { $correct++; } } echo 'Score: ' . (100 * $correct / count($evalData)) . '%' . PHP_EOL; ``` ```ruby Ruby client = Anthropic::Client.new def content_text(message) message.content.filter_map { |block| block.text if block.type == :text }.join end def build_grader_prompt(answer, rubric) <<~PROMPT Grade this answer based on the rubric: #{rubric} #{answer} Think through your reasoning in tags, then output 'correct' or 'incorrect' in tags. PROMPT end def grade_completion(client, output, golden_answer) grader_response = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 2048, messages: [ { role: "user", content: build_grader_prompt(output, golden_answer) } ] ) content_text(grader_response).downcase.include?("correct") ? "correct" : "incorrect" end # Example usage eval_data = [ { question: "Is 42 the answer to life, the universe, and everything?", golden_answer: "Yes, according to 'The Hitchhiker's Guide to the Galaxy'." }, { question: "What is the capital of France?", golden_answer: "The capital of France is Paris." } ] def get_completion(client, prompt) message = client.messages.create( model: Anthropic::Model::CLAUDE_OPUS_5, max_tokens: 1024, messages: [ { role: "user", content: prompt } ] ) content_text(message) end grades = eval_data.map do |item| output = get_completion(client, item[:question]) grade_completion(client, output, item[:golden_answer]) end puts "Score: #{100.0 * grades.count("correct") / grades.length}%" ``` ## Next steps Brainstorm success criteria for your use case with Claude on claude.ai.\ \ **Tip:** Drop this page into the chat as guidance for Claude! More code examples of human-, code-, and LLM-graded evals.