--- title: LLM Evaluation (Evals) description: LLM evaluation measures whether a language-model application actually works. You will learn error analysis, building test sets, code-based and model-graded checks, benchmarks and their limits, and how to monitor quality once a system is in production. category: programming-tech subcategory: artificial-intelligence difficulty: beginner, intermediate, advanced url: /subject/llm-evaluation --- # LLM Evaluation (Evals) LLM evaluation measures whether a language-model application actually works. You will learn error analysis, building test sets, code-based and model-graded checks, benchmarks and their limits, and how to monitor quality once a system is in production. ## Available Resources 1 Books • 1 Courses • 5 Websites • 2 Papers ## Courses ### 1. Improving Accuracy of LLM Applications **Author:** Sharon Zhou, Amit Sangani Walks through a systematic loop for making LLM applications reliable: build evaluation metrics, apply prompting and self-reflection, then fine-tune with LoRA and memory tuning. The running example is a text-to-SQL agent that hallucinates less each iteration. **Difficulty:** Intermediate | **Language:** English | **Price:** Free **Link:** https://www.deeplearning.ai/courses/improving-accuracy-of-llm-applications **Tags:** llm-evaluation, fine-tuning, lora, hallucination, text-to-sql ## Websites ### 1. Inspect: Open-Source Framework for LLM Evaluations **Author:** UK AI Security Institute Documentation and tutorial for Inspect, the UK AI Security Institute's open-source Python evaluation framework. Covers datasets, solvers, custom scorers, model grading, sandboxed agent evals, log analysis, and running evaluation sets across models, with over 200 prebuilt evals. **Difficulty:** Intermediate | **Language:** English | **Price:** Free **Link:** https://inspect.aisi.org.uk/ **Tags:** llm-evaluation, inspect-ai, agent-evals, model-grading, python ### 2. Evaluating the Effectiveness of LLM-Evaluators (aka LLM-as-Judge) **Author:** Eugene Yan Literature review of about two dozen papers on using LLMs to grade LLM output. Covers use cases, prompting techniques, aligning judges to human labels, fine-tuned evaluator models, and documented weaknesses such as position, verbosity, and self-preference bias. The standard deep reference on model-graded checks, the most error-prone part of any eval stack. **Difficulty:** Advanced | **Language:** English | **Price:** Free **Link:** https://eugeneyan.com/writing/llm-evaluators/ **Tags:** llm-as-judge, llm-evaluation, evaluator-bias, human-alignment, literature-review ### 3. Demystifying Evals for AI Agents **Author:** Mikaela Grace, Jeremy Hadfield, Rodrigo Olivares, Jiri De Jonghe Anthropic engineering guide to evaluating multi-step, tool-using agents. Defines tasks, trials, transcripts, and outcomes. Compares code-based, model-based, and human graders, explains pass@k versus pass^k, and separates capability evals from regression evals. Anthropic's newest and most substantive public writing on evals, and the clearest vocabulary for agent evaluation, where most new eval work now happens. It complements Hamel's product-centric material with grader taxonomy and trial statistics. **Difficulty:** Intermediate | **Language:** English | **Price:** Free **Link:** https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents **Tags:** agent-evals, llm-evaluation, ai-agents, graders, pass-at-k ### 4. AI Evals: Everything You Need to Know (LLM Evals FAQ) **Author:** Hamel Husain, Shreya Shankar Long question-and-answer guide drawn from the authors' evals course. Covers error analysis, synthetic test data, binary versus Likert grading, building LLM judges, human annotation, custom review tools, CI versus production monitoring, and guardrails versus evaluators. The densest free operational reference in the field, and awesome-evals lists it as a must-read. It covers the full topic scope: error analysis, test sets, code and model graders, and production monitoring. Most of the paid Maven course's methodology is here for free, which is why the course was discarded. **Difficulty:** Intermediate | **Language:** English | **Price:** Free **Link:** https://hamel.dev/blog/posts/evals-faq/ **Tags:** llm-evaluation, error-analysis, llm-as-judge, synthetic-data, production-monitoring, guardrails ### 5. Your AI Product Needs Evals **Author:** Hamel Husain Hamel Husain's 2024 essay arguing that LLM products improve through a loop of evaluating, debugging, and changing the system. Lays out three levels of evaluation: assertion-style unit tests, human and model review of logged traces, and A/B testing. It's short and free, and it frames the whole topic before readers move to the denser FAQ. **Difficulty:** Beginner | **Language:** English | **Price:** Free **Link:** https://hamel.dev/blog/posts/evals/ **Tags:** llm-evaluation, unit-testing, trace-review, ab-testing, llm-products ## Books ### 1. AI Engineering: Building Applications with Foundation Models **Author:** Chip Huyen Chip Huyen's 2025 O'Reilly book on building foundation-model applications. Chapters 3 and 4 cover evaluation: language-modeling metrics, exact evaluation, AI as a judge, comparative ranking, evaluation criteria, model selection, and designing an evaluation pipeline. The most rigorous textbook treatment of LLM evaluation methodology. Its two evaluation chapters (about 95 pages) give the conceptual grounding that blog posts skip. **Difficulty:** Intermediate | **Language:** English | **Price:** Paid **Link:** https://www.amazon.com/dp/1098166302?tag=edmonddante07-20 **Tags:** ai-engineering, foundation-models, llm-evaluation, llm-as-judge, model-selection ## Papers ### 1. Holistic Evaluation of Language Models (HELM) **Author:** Percy Liang, Rishi Bommasani, Tony Lee, et al. Stanford CRFM benchmark paper that evaluates 30 language models across 42 scenarios and seven metrics: accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency. Shows how benchmark coverage and standardized conditions change what model comparisons mean. HELM is the canonical argument that single-number leaderboards mislead and that evaluation needs defined scenarios and multiple metrics. **Difficulty:** Advanced | **Language:** English | **Price:** Free **Link:** https://arxiv.org/abs/2211.09110 **Tags:** llm-evaluation, benchmarks, helm, calibration, robustness ### 2. Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences **Author:** Shreya Shankar, J.D. Zamfirescu-Pereira, Björn Hartmann, Aditya G. Parameswaran, Ian Arawjo UIST 2024 paper introducing EvalGen, an interface that helps people align LLM-generated graders with their own judgments. Its user study identifies criteria drift: people refine their grading criteria as they grade outputs, so criteria cannot be fixed in advance. **Difficulty:** Advanced | **Language:** English | **Price:** Free **Link:** https://arxiv.org/abs/2404.12272 **Tags:** llm-as-judge, llm-evaluation, human-alignment, criteria-drift, evalgen --- *This content is part of Dantes.io - Your Treasure Map to Knowledge* *Curated by humans at Dantes.io. Personal study use welcome; republishing this curation requires permission (team@dantes.io).* View this page online: https://dantes.io/subject/llm-evaluation