[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M70K8FYCS5B7TM4PWZC8DWFS":3,"skill-same:01M70K8FYCS5B7TM4PWZC8DWFS":71},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":42,"linkOnly":43,"markdownContent":44,"markdownBody":45,"frontMatter":46,"refName":47,"commitSha":48,"skillDirectory":49,"forksCount":50,"repositoryHtmlUrl":51,"repositoryHomepageUrl":52,"ownerHtmlUrl":53,"ownerBlogUrl":54,"ownerDescription":55,"updatedAt":56,"install":57,"resources":70},"01M70K8FYCS5B7TM4PWZC8DWFS","llm-evaluation",null,"Implement comprehensive evaluation strategies for LLM applications using automated metrics, human feedback, and benchmarking. Use when testing LLM performance, measuring AI application quality, or establishing evaluation frameworks.","shengdabai\u002FTony-Claude-Code-Skills","shengdabai","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F116417419?v=4","01M70F6D2WY5PQV5A92BKYJQEW","Tony Sheng","01M70F6D3GAAAYZ5H4CC1G8VXM",3,"MIT","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002F_collections\u002Fwshobson-agents\u002Fplugins\u002Fllm-application-dev\u002Fskills\u002Fllm-evaluation\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41],"llm evaluation","Evaluation","application","quality","implement","use","applications","frameworks","for","feedback","metrics","human","benchmarking","strategies","LLM Applications","when","testing","performance","2026-11-10 10:12:30",false,"---\nname: llm-evaluation\ndescription: Implement comprehensive evaluation strategies for LLM applications using automated metrics, human feedback, and benchmarking. Use when testing LLM performance, measuring AI application quality, or establishing evaluation frameworks.\n---\n\n# LLM Evaluation\n\nMaster comprehensive evaluation strategies for LLM applications, from automated metrics to human evaluation and A\u002FB testing.\n\n## When to Use This Skill\n\n- Measuring LLM application performance systematically\n- Comparing different models or prompts\n- Detecting performance regressions before deployment\n- Validating improvements from prompt changes\n- Building confidence in production systems\n- Establishing baselines and tracking progress over time\n- Debugging unexpected model behavior\n\n## Core Evaluation Types\n\n### 1. Automated Metrics\nFast, repeatable, scalable evaluation using computed scores.\n\n**Text Generation:**\n- **BLEU**: N-gram overlap (translation)\n- **ROUGE**: Recall-oriented (summarization)\n- **METEOR**: Semantic similarity\n- **BERTScore**: Embedding-based similarity\n- **Perplexity**: Language model confidence\n\n**Classification:**\n- **Accuracy**: Percentage correct\n- **Precision\u002FRecall\u002FF1**: Class-specific performance\n- **Confusion Matrix**: Error patterns\n- **AUC-ROC**: Ranking quality\n\n**Retrieval (RAG):**\n- **MRR**: Mean Reciprocal Rank\n- **NDCG**: Normalized Discounted Cumulative Gain\n- **Precision@K**: Relevant in top K\n- **Recall@K**: Coverage in top K\n\n### 2. Human Evaluation\nManual assessment for quality aspects difficult to automate.\n\n**Dimensions:**\n- **Accuracy**: Factual correctness\n- **Coherence**: Logical flow\n- **Relevance**: Answers the question\n- **Fluency**: Natural language quality\n- **Safety**: No harmful content\n- **Helpfulness**: Useful to the user\n\n### 3. LLM-as-Judge\nUse stronger LLMs to evaluate weaker model outputs.\n\n**Approaches:**\n- **Pointwise**: Score individual responses\n- **Pairwise**: Compare two responses\n- **Reference-based**: Compare to gold standard\n- **Reference-free**: Judge without ground truth\n\n## Quick Start\n\n```python\nfrom llm_eval import EvaluationSuite, Metric\n\n# Define evaluation suite\nsuite = EvaluationSuite([\n    Metric.accuracy(),\n    Metric.bleu(),\n    Metric.bertscore(),\n    Metric.custom(name=\"groundedness\", fn=check_groundedness)\n])\n\n# Prepare test cases\ntest_cases = [\n    {\n        \"input\": \"What is the capital of France?\",\n        \"expected\": \"Paris\",\n        \"context\": \"France is a country in Europe. Paris is its capital.\"\n    },\n    # ... more test cases\n]\n\n# Run evaluation\nresults = suite.evaluate(\n    model=your_model,\n    test_cases=test_cases\n)\n\nprint(f\"Overall Accuracy: {results.metrics['accuracy']}\")\nprint(f\"BLEU Score: {results.metrics['bleu']}\")\n```\n\n## Automated Metrics Implementation\n\n### BLEU Score\n```python\nfrom nltk.translate.bleu_score import sentence_bleu, SmoothingFunction\n\ndef calculate_bleu(reference, hypothesis):\n    \"\"\"Calculate BLEU score between reference and hypothesis.\"\"\"\n    smoothie = SmoothingFunction().method4\n\n    return sentence_bleu(\n        [reference.split()],\n        hypothesis.split(),\n        smoothing_function=smoothie\n    )\n\n# Usage\nbleu = calculate_bleu(\n    reference=\"The cat sat on the mat\",\n    hypothesis=\"A cat is sitting on the mat\"\n)\n```\n\n### ROUGE Score\n```python\nfrom rouge_score import rouge_scorer\n\ndef calculate_rouge(reference, hypothesis):\n    \"\"\"Calculate ROUGE scores.\"\"\"\n    scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)\n    scores = scorer.score(reference, hypothesis)\n\n    return {\n        'rouge1': scores['rouge1'].fmeasure,\n        'rouge2': scores['rouge2'].fmeasure,\n        'rougeL': scores['rougeL'].fmeasure\n    }\n```\n\n### BERTScore\n```python\nfrom bert_score import score\n\ndef calculate_bertscore(references, hypotheses):\n    \"\"\"Calculate BERTScore using pre-trained BERT.\"\"\"\n    P, R, F1 = score(\n        hypotheses,\n        references,\n        lang='en',\n        model_type='microsoft\u002Fdeberta-xlarge-mnli'\n    )\n\n    return {\n        'precision': P.mean().item(),\n        'recall': R.mean().item(),\n        'f1': F1.mean().item()\n    }\n```\n\n### Custom Metrics\n```python\ndef calculate_groundedness(response, context):\n    \"\"\"Check if response is grounded in provided context.\"\"\"\n    # Use NLI model to check entailment\n    from transformers import pipeline\n\n    nli = pipeline(\"text-classification\", model=\"microsoft\u002Fdeberta-large-mnli\")\n\n    result = nli(f\"{context} [SEP] {response}\")[0]\n\n    # Return confidence that response is entailed by context\n    return result['score'] if result['label'] == 'ENTAILMENT' else 0.0\n\ndef calculate_toxicity(text):\n    \"\"\"Measure toxicity in generated text.\"\"\"\n    from detoxify import Detoxify\n\n    results = Detoxify('original').predict(text)\n    return max(results.values())  # Return highest toxicity score\n\ndef calculate_factuality(claim, knowledge_base):\n    \"\"\"Verify factual claims against knowledge base.\"\"\"\n    # Implementation depends on your knowledge base\n    # Could use retrieval + NLI, or fact-checking API\n    pass\n```\n\n## LLM-as-Judge Patterns\n\n### Single Output Evaluation\n```python\ndef llm_judge_quality(response, question):\n    \"\"\"Use GPT-5 to judge response quality.\"\"\"\n    prompt = f\"\"\"Rate the following response on a scale of 1-10 for:\n1. Accuracy (factually correct)\n2. Helpfulness (answers the question)\n3. Clarity (well-written and understandable)\n\nQuestion: {question}\nResponse: {response}\n\nProvide ratings in JSON format:\n{{\n  \"accuracy\": \u003C1-10>,\n  \"helpfulness\": \u003C1-10>,\n  \"clarity\": \u003C1-10>,\n  \"reasoning\": \"\u003Cbrief explanation>\"\n}}\n\"\"\"\n\n    result = openai.ChatCompletion.create(\n        model=\"gpt-5\",\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n        temperature=0\n    )\n\n    return json.loads(result.choices[0].message.content)\n```\n\n### Pairwise Comparison\n```python\ndef compare_responses(question, response_a, response_b):\n    \"\"\"Compare two responses using LLM judge.\"\"\"\n    prompt = f\"\"\"Compare these two responses to the question and determine which is better.\n\nQuestion: {question}\n\nResponse A: {response_a}\n\nResponse B: {response_b}\n\nWhich response is better and why? Consider accuracy, helpfulness, and clarity.\n\nAnswer with JSON:\n{{\n  \"winner\": \"A\" or \"B\" or \"tie\",\n  \"reasoning\": \"\u003Cexplanation>\",\n  \"confidence\": \u003C1-10>\n}}\n\"\"\"\n\n    result = openai.ChatCompletion.create(\n        model=\"gpt-5\",\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n        temperature=0\n    )\n\n    return json.loads(result.choices[0].message.content)\n```\n\n## Human Evaluation Frameworks\n\n### Annotation Guidelines\n```python\nclass AnnotationTask:\n    \"\"\"Structure for human annotation task.\"\"\"\n\n    def __init__(self, response, question, context=None):\n        self.response = response\n        self.question = question\n        self.context = context\n\n    def get_annotation_form(self):\n        return {\n            \"question\": self.question,\n            \"context\": self.context,\n            \"response\": self.response,\n            \"ratings\": {\n                \"accuracy\": {\n                    \"scale\": \"1-5\",\n                    \"description\": \"Is the response factually correct?\"\n                },\n                \"relevance\": {\n                    \"scale\": \"1-5\",\n                    \"description\": \"Does it answer the question?\"\n                },\n                \"coherence\": {\n                    \"scale\": \"1-5\",\n                    \"description\": \"Is it logically consistent?\"\n                }\n            },\n            \"issues\": {\n                \"factual_error\": False,\n                \"hallucination\": False,\n                \"off_topic\": False,\n                \"unsafe_content\": False\n            },\n            \"feedback\": \"\"\n        }\n```\n\n### Inter-Rater Agreement\n```python\nfrom sklearn.metrics import cohen_kappa_score\n\ndef calculate_agreement(rater1_scores, rater2_scores):\n    \"\"\"Calculate inter-rater agreement.\"\"\"\n    kappa = cohen_kappa_score(rater1_scores, rater2_scores)\n\n    interpretation = {\n        kappa \u003C 0: \"Poor\",\n        kappa \u003C 0.2: \"Slight\",\n        kappa \u003C 0.4: \"Fair\",\n        kappa \u003C 0.6: \"Moderate\",\n        kappa \u003C 0.8: \"Substantial\",\n        kappa \u003C= 1.0: \"Almost Perfect\"\n    }\n\n    return {\n        \"kappa\": kappa,\n        \"interpretation\": interpretation[True]\n    }\n```\n\n## A\u002FB Testing\n\n### Statistical Testing Framework\n```python\nfrom scipy import stats\nimport numpy as np\n\nclass ABTest:\n    def __init__(self, variant_a_name=\"A\", variant_b_name=\"B\"):\n        self.variant_a = {\"name\": variant_a_name, \"scores\": []}\n        self.variant_b = {\"name\": variant_b_name, \"scores\": []}\n\n    def add_result(self, variant, score):\n        \"\"\"Add evaluation result for a variant.\"\"\"\n        if variant == \"A\":\n            self.variant_a[\"scores\"].append(score)\n        else:\n            self.variant_b[\"scores\"].append(score)\n\n    def analyze(self, alpha=0.05):\n        \"\"\"Perform statistical analysis.\"\"\"\n        a_scores = self.variant_a[\"scores\"]\n        b_scores = self.variant_b[\"scores\"]\n\n        # T-test\n        t_stat, p_value = stats.ttest_ind(a_scores, b_scores)\n\n        # Effect size (Cohen's d)\n        pooled_std = np.sqrt((np.std(a_scores)**2 + np.std(b_scores)**2) \u002F 2)\n        cohens_d = (np.mean(b_scores) - np.mean(a_scores)) \u002F pooled_std\n\n        return {\n            \"variant_a_mean\": np.mean(a_scores),\n            \"variant_b_mean\": np.mean(b_scores),\n            \"difference\": np.mean(b_scores) - np.mean(a_scores),\n            \"relative_improvement\": (np.mean(b_scores) - np.mean(a_scores)) \u002F np.mean(a_scores),\n            \"p_value\": p_value,\n            \"statistically_significant\": p_value \u003C alpha,\n            \"cohens_d\": cohens_d,\n            \"effect_size\": self.interpret_cohens_d(cohens_d),\n            \"winner\": \"B\" if np.mean(b_scores) > np.mean(a_scores) else \"A\"\n        }\n\n    @staticmethod\n    def interpret_cohens_d(d):\n        \"\"\"Interpret Cohen's d effect size.\"\"\"\n        abs_d = abs(d)\n        if abs_d \u003C 0.2:\n            return \"negligible\"\n        elif abs_d \u003C 0.5:\n            return \"small\"\n        elif abs_d \u003C 0.8:\n            return \"medium\"\n        else:\n            return \"large\"\n```\n\n## Regression Testing\n\n### Regression Detection\n```python\nclass RegressionDetector:\n    def __init__(self, baseline_results, threshold=0.05):\n        self.baseline = baseline_results\n        self.threshold = threshold\n\n    def check_for_regression(self, new_results):\n        \"\"\"Detect if new results show regression.\"\"\"\n        regressions = []\n\n        for metric in self.baseline.keys():\n            baseline_score = self.baseline[metric]\n            new_score = new_results.get(metric)\n\n            if new_score is None:\n                continue\n\n            # Calculate relative change\n            relative_change = (new_score - baseline_score) \u002F baseline_score\n\n            # Flag if significant decrease\n            if relative_change \u003C -self.threshold:\n                regressions.append({\n                    \"metric\": metric,\n                    \"baseline\": baseline_score,\n                    \"current\": new_score,\n                    \"change\": relative_change\n                })\n\n        return {\n            \"has_regression\": len(regressions) > 0,\n            \"regressions\": regressions\n        }\n```\n\n## Benchmarking\n\n### Running Benchmarks\n```python\nclass BenchmarkRunner:\n    def __init__(self, benchmark_dataset):\n        self.dataset = benchmark_dataset\n\n    def run_benchmark(self, model, metrics):\n        \"\"\"Run model on benchmark and calculate metrics.\"\"\"\n        results = {metric.name: [] for metric in metrics}\n\n        for example in self.dataset:\n            # Generate prediction\n            prediction = model.predict(example[\"input\"])\n\n            # Calculate each metric\n            for metric in metrics:\n                score = metric.calculate(\n                    prediction=prediction,\n                    reference=example[\"reference\"],\n                    context=example.get(\"context\")\n                )\n                results[metric.name].append(score)\n\n        # Aggregate results\n        return {\n            metric: {\n                \"mean\": np.mean(scores),\n                \"std\": np.std(scores),\n                \"min\": min(scores),\n                \"max\": max(scores)\n            }\n            for metric, scores in results.items()\n        }\n```\n\n## Resources\n\n- **references\u002Fmetrics.md**: Comprehensive metric guide\n- **references\u002Fhuman-evaluation.md**: Annotation best practices\n- **references\u002Fbenchmarking.md**: Standard benchmarks\n- **references\u002Fa-b-testing.md**: Statistical testing guide\n- **references\u002Fregression-testing.md**: CI\u002FCD integration\n- **assets\u002Fevaluation-framework.py**: Complete evaluation harness\n- **assets\u002Fbenchmark-dataset.jsonl**: Example datasets\n- **scripts\u002Fevaluate-model.py**: Automated evaluation runner\n\n## Best Practices\n\n1. **Multiple Metrics**: Use diverse metrics for comprehensive view\n2. **Representative Data**: Test on real-world, diverse examples\n3. **Baselines**: Always compare against baseline performance\n4. **Statistical Rigor**: Use proper statistical tests for comparisons\n5. **Continuous Evaluation**: Integrate into CI\u002FCD pipeline\n6. **Human Validation**: Combine automated metrics with human judgment\n7. **Error Analysis**: Investigate failures to understand weaknesses\n8. **Version Control**: Track evaluation results over time\n\n## Common Pitfalls\n\n- **Single Metric Obsession**: Optimizing for one metric at the expense of others\n- **Small Sample Size**: Drawing conclusions from too few examples\n- **Data Contamination**: Testing on training data\n- **Ignoring Variance**: Not accounting for statistical uncertainty\n- **Metric Mismatch**: Using metrics not aligned with business goals\n","# LLM Evaluation\n\nMaster comprehensive evaluation strategies for LLM applications, from automated metrics to human evaluation and A\u002FB testing.\n\n## When to Use This Skill\n\n- Measuring LLM application performance systematically\n- Comparing different models or prompts\n- Detecting performance regressions before deployment\n- Validating improvements from prompt changes\n- Building confidence in production systems\n- Establishing baselines and tracking progress over time\n- Debugging unexpected model behavior\n\n## Core Evaluation Types\n\n### 1. Automated Metrics\nFast, repeatable, scalable evaluation using computed scores.\n\n**Text Generation:**\n- **BLEU**: N-gram overlap (translation)\n- **ROUGE**: Recall-oriented (summarization)\n- **METEOR**: Semantic similarity\n- **BERTScore**: Embedding-based similarity\n- **Perplexity**: Language model confidence\n\n**Classification:**\n- **Accuracy**: Percentage correct\n- **Precision\u002FRecall\u002FF1**: Class-specific performance\n- **Confusion Matrix**: Error patterns\n- **AUC-ROC**: Ranking quality\n\n**Retrieval (RAG):**\n- **MRR**: Mean Reciprocal Rank\n- **NDCG**: Normalized Discounted Cumulative Gain\n- **Precision@K**: Relevant in top K\n- **Recall@K**: Coverage in top K\n\n### 2. Human Evaluation\nManual assessment for quality aspects difficult to automate.\n\n**Dimensions:**\n- **Accuracy**: Factual correctness\n- **Coherence**: Logical flow\n- **Relevance**: Answers the question\n- **Fluency**: Natural language quality\n- **Safety**: No harmful content\n- **Helpfulness**: Useful to the user\n\n### 3. LLM-as-Judge\nUse stronger LLMs to evaluate weaker model outputs.\n\n**Approaches:**\n- **Pointwise**: Score individual responses\n- **Pairwise**: Compare two responses\n- **Reference-based**: Compare to gold standard\n- **Reference-free**: Judge without ground truth\n\n## Quick Start\n\n```python\nfrom llm_eval import EvaluationSuite, Metric\n\n# Define evaluation suite\nsuite = EvaluationSuite([\n    Metric.accuracy(),\n    Metric.bleu(),\n    Metric.bertscore(),\n    Metric.custom(name=\"groundedness\", fn=check_groundedness)\n])\n\n# Prepare test cases\ntest_cases = [\n    {\n        \"input\": \"What is the capital of France?\",\n        \"expected\": \"Paris\",\n        \"context\": \"France is a country in Europe. Paris is its capital.\"\n    },\n    # ... more test cases\n]\n\n# Run evaluation\nresults = suite.evaluate(\n    model=your_model,\n    test_cases=test_cases\n)\n\nprint(f\"Overall Accuracy: {results.metrics['accuracy']}\")\nprint(f\"BLEU Score: {results.metrics['bleu']}\")\n```\n\n## Automated Metrics Implementation\n\n### BLEU Score\n```python\nfrom nltk.translate.bleu_score import sentence_bleu, SmoothingFunction\n\ndef calculate_bleu(reference, hypothesis):\n    \"\"\"Calculate BLEU score between reference and hypothesis.\"\"\"\n    smoothie = SmoothingFunction().method4\n\n    return sentence_bleu(\n        [reference.split()],\n        hypothesis.split(),\n        smoothing_function=smoothie\n    )\n\n# Usage\nbleu = calculate_bleu(\n    reference=\"The cat sat on the mat\",\n    hypothesis=\"A cat is sitting on the mat\"\n)\n```\n\n### ROUGE Score\n```python\nfrom rouge_score import rouge_scorer\n\ndef calculate_rouge(reference, hypothesis):\n    \"\"\"Calculate ROUGE scores.\"\"\"\n    scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)\n    scores = scorer.score(reference, hypothesis)\n\n    return {\n        'rouge1': scores['rouge1'].fmeasure,\n        'rouge2': scores['rouge2'].fmeasure,\n        'rougeL': scores['rougeL'].fmeasure\n    }\n```\n\n### BERTScore\n```python\nfrom bert_score import score\n\ndef calculate_bertscore(references, hypotheses):\n    \"\"\"Calculate BERTScore using pre-trained BERT.\"\"\"\n    P, R, F1 = score(\n        hypotheses,\n        references,\n        lang='en',\n        model_type='microsoft\u002Fdeberta-xlarge-mnli'\n    )\n\n    return {\n        'precision': P.mean().item(),\n        'recall': R.mean().item(),\n        'f1': F1.mean().item()\n    }\n```\n\n### Custom Metrics\n```python\ndef calculate_groundedness(response, context):\n    \"\"\"Check if response is grounded in provided context.\"\"\"\n    # Use NLI model to check entailment\n    from transformers import pipeline\n\n    nli = pipeline(\"text-classification\", model=\"microsoft\u002Fdeberta-large-mnli\")\n\n    result = nli(f\"{context} [SEP] {response}\")[0]\n\n    # Return confidence that response is entailed by context\n    return result['score'] if result['label'] == 'ENTAILMENT' else 0.0\n\ndef calculate_toxicity(text):\n    \"\"\"Measure toxicity in generated text.\"\"\"\n    from detoxify import Detoxify\n\n    results = Detoxify('original').predict(text)\n    return max(results.values())  # Return highest toxicity score\n\ndef calculate_factuality(claim, knowledge_base):\n    \"\"\"Verify factual claims against knowledge base.\"\"\"\n    # Implementation depends on your knowledge base\n    # Could use retrieval + NLI, or fact-checking API\n    pass\n```\n\n## LLM-as-Judge Patterns\n\n### Single Output Evaluation\n```python\ndef llm_judge_quality(response, question):\n    \"\"\"Use GPT-5 to judge response quality.\"\"\"\n    prompt = f\"\"\"Rate the following response on a scale of 1-10 for:\n1. Accuracy (factually correct)\n2. Helpfulness (answers the question)\n3. Clarity (well-written and understandable)\n\nQuestion: {question}\nResponse: {response}\n\nProvide ratings in JSON format:\n{{\n  \"accuracy\": \u003C1-10>,\n  \"helpfulness\": \u003C1-10>,\n  \"clarity\": \u003C1-10>,\n  \"reasoning\": \"\u003Cbrief explanation>\"\n}}\n\"\"\"\n\n    result = openai.ChatCompletion.create(\n        model=\"gpt-5\",\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n        temperature=0\n    )\n\n    return json.loads(result.choices[0].message.content)\n```\n\n### Pairwise Comparison\n```python\ndef compare_responses(question, response_a, response_b):\n    \"\"\"Compare two responses using LLM judge.\"\"\"\n    prompt = f\"\"\"Compare these two responses to the question and determine which is better.\n\nQuestion: {question}\n\nResponse A: {response_a}\n\nResponse B: {response_b}\n\nWhich response is better and why? Consider accuracy, helpfulness, and clarity.\n\nAnswer with JSON:\n{{\n  \"winner\": \"A\" or \"B\" or \"tie\",\n  \"reasoning\": \"\u003Cexplanation>\",\n  \"confidence\": \u003C1-10>\n}}\n\"\"\"\n\n    result = openai.ChatCompletion.create(\n        model=\"gpt-5\",\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n        temperature=0\n    )\n\n    return json.loads(result.choices[0].message.content)\n```\n\n## Human Evaluation Frameworks\n\n### Annotation Guidelines\n```python\nclass AnnotationTask:\n    \"\"\"Structure for human annotation task.\"\"\"\n\n    def __init__(self, response, question, context=None):\n        self.response = response\n        self.question = question\n        self.context = context\n\n    def get_annotation_form(self):\n        return {\n            \"question\": self.question,\n            \"context\": self.context,\n            \"response\": self.response,\n            \"ratings\": {\n                \"accuracy\": {\n                    \"scale\": \"1-5\",\n                    \"description\": \"Is the response factually correct?\"\n                },\n                \"relevance\": {\n                    \"scale\": \"1-5\",\n                    \"description\": \"Does it answer the question?\"\n                },\n                \"coherence\": {\n                    \"scale\": \"1-5\",\n                    \"description\": \"Is it logically consistent?\"\n                }\n            },\n            \"issues\": {\n                \"factual_error\": False,\n                \"hallucination\": False,\n                \"off_topic\": False,\n                \"unsafe_content\": False\n            },\n            \"feedback\": \"\"\n        }\n```\n\n### Inter-Rater Agreement\n```python\nfrom sklearn.metrics import cohen_kappa_score\n\ndef calculate_agreement(rater1_scores, rater2_scores):\n    \"\"\"Calculate inter-rater agreement.\"\"\"\n    kappa = cohen_kappa_score(rater1_scores, rater2_scores)\n\n    interpretation = {\n        kappa \u003C 0: \"Poor\",\n        kappa \u003C 0.2: \"Slight\",\n        kappa \u003C 0.4: \"Fair\",\n        kappa \u003C 0.6: \"Moderate\",\n        kappa \u003C 0.8: \"Substantial\",\n        kappa \u003C= 1.0: \"Almost Perfect\"\n    }\n\n    return {\n        \"kappa\": kappa,\n        \"interpretation\": interpretation[True]\n    }\n```\n\n## A\u002FB Testing\n\n### Statistical Testing Framework\n```python\nfrom scipy import stats\nimport numpy as np\n\nclass ABTest:\n    def __init__(self, variant_a_name=\"A\", variant_b_name=\"B\"):\n        self.variant_a = {\"name\": variant_a_name, \"scores\": []}\n        self.variant_b = {\"name\": variant_b_name, \"scores\": []}\n\n    def add_result(self, variant, score):\n        \"\"\"Add evaluation result for a variant.\"\"\"\n        if variant == \"A\":\n            self.variant_a[\"scores\"].append(score)\n        else:\n            self.variant_b[\"scores\"].append(score)\n\n    def analyze(self, alpha=0.05):\n        \"\"\"Perform statistical analysis.\"\"\"\n        a_scores = self.variant_a[\"scores\"]\n        b_scores = self.variant_b[\"scores\"]\n\n        # T-test\n        t_stat, p_value = stats.ttest_ind(a_scores, b_scores)\n\n        # Effect size (Cohen's d)\n        pooled_std = np.sqrt((np.std(a_scores)**2 + np.std(b_scores)**2) \u002F 2)\n        cohens_d = (np.mean(b_scores) - np.mean(a_scores)) \u002F pooled_std\n\n        return {\n            \"variant_a_mean\": np.mean(a_scores),\n            \"variant_b_mean\": np.mean(b_scores),\n            \"difference\": np.mean(b_scores) - np.mean(a_scores),\n            \"relative_improvement\": (np.mean(b_scores) - np.mean(a_scores)) \u002F np.mean(a_scores),\n            \"p_value\": p_value,\n            \"statistically_significant\": p_value \u003C alpha,\n            \"cohens_d\": cohens_d,\n            \"effect_size\": self.interpret_cohens_d(cohens_d),\n            \"winner\": \"B\" if np.mean(b_scores) > np.mean(a_scores) else \"A\"\n        }\n\n    @staticmethod\n    def interpret_cohens_d(d):\n        \"\"\"Interpret Cohen's d effect size.\"\"\"\n        abs_d = abs(d)\n        if abs_d \u003C 0.2:\n            return \"negligible\"\n        elif abs_d \u003C 0.5:\n            return \"small\"\n        elif abs_d \u003C 0.8:\n            return \"medium\"\n        else:\n            return \"large\"\n```\n\n## Regression Testing\n\n### Regression Detection\n```python\nclass RegressionDetector:\n    def __init__(self, baseline_results, threshold=0.05):\n        self.baseline = baseline_results\n        self.threshold = threshold\n\n    def check_for_regression(self, new_results):\n        \"\"\"Detect if new results show regression.\"\"\"\n        regressions = []\n\n        for metric in self.baseline.keys():\n            baseline_score = self.baseline[metric]\n            new_score = new_results.get(metric)\n\n            if new_score is None:\n                continue\n\n            # Calculate relative change\n            relative_change = (new_score - baseline_score) \u002F baseline_score\n\n            # Flag if significant decrease\n            if relative_change \u003C -self.threshold:\n                regressions.append({\n                    \"metric\": metric,\n                    \"baseline\": baseline_score,\n                    \"current\": new_score,\n                    \"change\": relative_change\n                })\n\n        return {\n            \"has_regression\": len(regressions) > 0,\n            \"regressions\": regressions\n        }\n```\n\n## Benchmarking\n\n### Running Benchmarks\n```python\nclass BenchmarkRunner:\n    def __init__(self, benchmark_dataset):\n        self.dataset = benchmark_dataset\n\n    def run_benchmark(self, model, metrics):\n        \"\"\"Run model on benchmark and calculate metrics.\"\"\"\n        results = {metric.name: [] for metric in metrics}\n\n        for example in self.dataset:\n            # Generate prediction\n            prediction = model.predict(example[\"input\"])\n\n            # Calculate each metric\n            for metric in metrics:\n                score = metric.calculate(\n                    prediction=prediction,\n                    reference=example[\"reference\"],\n                    context=example.get(\"context\")\n                )\n                results[metric.name].append(score)\n\n        # Aggregate results\n        return {\n            metric: {\n                \"mean\": np.mean(scores),\n                \"std\": np.std(scores),\n                \"min\": min(scores),\n                \"max\": max(scores)\n            }\n            for metric, scores in results.items()\n        }\n```\n\n## Resources\n\n- **references\u002Fmetrics.md**: Comprehensive metric guide\n- **references\u002Fhuman-evaluation.md**: Annotation best practices\n- **references\u002Fbenchmarking.md**: Standard benchmarks\n- **references\u002Fa-b-testing.md**: Statistical testing guide\n- **references\u002Fregression-testing.md**: CI\u002FCD integration\n- **assets\u002Fevaluation-framework.py**: Complete evaluation harness\n- **assets\u002Fbenchmark-dataset.jsonl**: Example datasets\n- **scripts\u002Fevaluate-model.py**: Automated evaluation runner\n\n## Best Practices\n\n1. **Multiple Metrics**: Use diverse metrics for comprehensive view\n2. **Representative Data**: Test on real-world, diverse examples\n3. **Baselines**: Always compare against baseline performance\n4. **Statistical Rigor**: Use proper statistical tests for comparisons\n5. **Continuous Evaluation**: Integrate into CI\u002FCD pipeline\n6. **Human Validation**: Combine automated metrics with human judgment\n7. **Error Analysis**: Investigate failures to understand weaknesses\n8. **Version Control**: Track evaluation results over time\n\n## Common Pitfalls\n\n- **Single Metric Obsession**: Optimizing for one metric at the expense of others\n- **Small Sample Size**: Drawing conclusions from too few examples\n- **Data Contamination**: Testing on training data\n- **Ignoring Variance**: Not accounting for statistical uncertainty\n- **Metric Mismatch**: Using metrics not aligned with business goals\n",{"name":5,"description":7},"main","7850740bd837adb32f0a109ab2af82482196c6f2","skills\u002F_collections\u002Fwshobson-agents\u002Fplugins\u002Fllm-application-dev\u002Fskills\u002Fllm-evaluation",1,"https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills","https:\u002F\u002Fzturnsgo.com","https:\u002F\u002Fgithub.com\u002Fshengdabai","https:\u002F\u002Fzturnsgo.com\u002F","AI × education creator — producer, not consumer. Taught 6,000+ learners; shipping learning tools & daily essays with Claude Code, practicing @xiaolai's way 🥋","2026-11-01 21:28:23",{"source":8,"skillName":5,"commands":58,"prompt":68,"zipAvailable":69},[59,62,65],{"tool":60,"command":61},"npx","npx skills add shengdabai\u002FTony-Claude-Code-Skills --skill llm-evaluation",{"tool":63,"command":64},"bunx","bunx skills add shengdabai\u002FTony-Claude-Code-Skills --skill llm-evaluation",{"tool":66,"command":67},"pnpm","pnpm dlx skills add shengdabai\u002FTony-Claude-Code-Skills --skill llm-evaluation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002F_collections\u002Fwshobson-agents\u002Fplugins\u002Fllm-application-dev\u002Fskills\u002Fllm-evaluation\u002FSKILL.md",true,[],[72,81,89,97,105,113],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":76,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":80},"01M73H7XKBHE1ETQDZT94MF9HH","seedance-prompt-en","Write effective prompts for Jimeng Seedance 2.0 multimodal AI video generation. Use when users want to create video prompts using text, images, videos, and audio inputs with the @ reference system. Covers camera movements, effects replication, video extension, editing, music beat-matching, e-commerce ads, short dramas, and educational content.","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fseedance-prompt-en\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fseedance-prompt-en\u002FSKILL.md",[],[],"2026-11-11 15:11:14",{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":85,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":80},"01M73GRE9998Q1DZWC71N41JWG","onepage-pdf","Convert an HTML page into a single continuous-page PDF (one tall page, no pagination breaks), preserving desktop layout, backgrounds and selectable text. Use when the user asks to turn an HTML file\u002Freport\u002Fproposal into a \"single-page PDF\", \"long PDF\", \"不分页 PDF\", \"单页 PDF\", \"长图式 PDF\", or complains that an HTML-to-PDF export breaks into pages or cuts content. Supports optional string redaction with leak verification.","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fonepage-pdf\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fonepage-pdf\u002FSKILL.md",[],[],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":93,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":80},"01M73GQG1KZDADVPV3MHAT1J5A","native-feel-cross-platform-desktop","Use when the user is designing, prototyping, or rewriting a desktop app that must run on multiple OSes (macOS + Windows, optionally Linux) AND feel indistinguishable from a native app to its users — fast launch, native windowing, native input handling, native materials. Trigger words include \"cross-platform desktop\", \"Electron alternative\", \"Tauri vs native\", \"WebView wrapper\", \"near-native performance\", \"Raycast architecture\", \"WebKit\u002FWebView2 quirks\", \"WKWebView\", \"system tray app\", \"global hotkey app\", \"launcher app\". Do NOT trigger this skill for pure web apps, pure mobile apps, or for greenfield projects that have no native-feel requirement.","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fnative-feel-cross-platform-desktop\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fnative-feel-cross-platform-desktop\u002FSKILL.md",[],[],{"skillNo":98,"name":99,"nameZh":6,"description":100,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":101,"sourceHtmlUrl":102,"categories":103,"keywords":104,"publishedAt":80},"01M73G87PDNSMKG2FHWCYER6N8","gstack","Router for the gstack skill suite. (gstack)","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fgstack\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fgstack\u002FSKILL.md",[],[],{"skillNo":106,"name":107,"nameZh":6,"description":108,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":109,"sourceHtmlUrl":110,"categories":111,"keywords":112,"publishedAt":80},"01M73G3EAQWKCM8DB8D0N0QSVG","fireworks-tech-graph","Use when the user wants to create any technical diagram - architecture, data flow, flowchart, sequence, agent\u002Fmemory, or concept map - and export as SVG+PNG. Trigger on: \"画图\" \"帮我画\" \"生成图\" \"做个图\" \"架构图\" \"流程图\" \"可视化一下\" \"出图\" \"generate diagram\" \"draw diagram\" \"visualize\" or any system\u002Fflow description the user wants illustrated.","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Ffireworks-tech-graph\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Ffireworks-tech-graph\u002FSKILL.md",[],[],{"skillNo":114,"name":115,"nameZh":6,"description":116,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":117,"categories":118,"keywords":119,"publishedAt":80},"01M73G086AXAH9DW1H5G5N3CPB","dbs-video-extract","接收抖音、小红书、微信视频号等短视频链接或分享文案，查询作品／账号数据并提取语音文字稿，按作者和标题归档为 Markdown。用户希望完整解析短视频、取得数据或文字稿时使用；运行前检查 TikHub 与轻抖使用凭证，缺少时用新手能理解的方式引导购买和安全配置。","https:\u002F\u002Fgithub.com\u002Fshengdabai\u002FTony-Claude-Code-Skills\u002Fblob\u002Fmain\u002Fskills\u002Fdbs-video-extract\u002FSKILL.md",[],[]]