Module 2: Benchmark and Harness › Benchmark Questions
GitHub Login Glossary

Designing Good Benchmark Questions

In the previous lesson, you wrote your first 10 benchmark questions. Some of them are probably good. Some of them probably aren't, and right now you might not be able to tell which is which. That's perfectly normal. Writing diagnostic questions is a skill in itself, and takes practice. In this lesson, we'll learn to improve our questions.

By the end, you'll have a better idea of what makes a question useful for measuring your system, what makes a question misleading, and how to grade the answers consistently.

What you'll learn

  • Distinguish diagnostic questions from vague or untestable ones
  • Apply a grading rubric (fully correct / partially correct / unsupported / wrong) consistently
  • Identify and fix common benchmark design mistakes
  • Write gold answers that are specific enough to grade against
  • Expand your benchmark set to 30+ questions with confidence

Concepts

Diagnostic question: a benchmark question where the answer clearly reveals whether the system has a specific capability. "Tell me about the codebase" is not diagnostic. Any answer could be partially right. "What function handles JWT token verification, and what happens when the token is expired?" is diagnostic. The answer is either correct and specific, or it isn't.

Grading rubric: a consistent set of criteria for evaluating answers. Without a rubric, grading is subjective and varies between sessions. The rubric we'll use has four levels:

GradeMeaning
Fully correctThe answer addresses the question accurately, is supported by the code, and doesn't contain fabricated details
Partially correctThe answer is on the right track but misses key details, includes minor inaccuracies, or is incomplete
UnsupportedThe answer sounds plausible but can't be verified from the codebase. It may be hallucinated
WrongThe answer is factually incorrect based on what the code actually does

Failure label: a tag that explains why a wrong or unsupported answer failed. Common labels: missing_evidence (the system lacked the information it needed (use this for the baseline before retrieval exists)retrieval_miss (the retrieval system didn't find the right code), wrong_chunk (it found related but wrong code), hallucination (it fabricated details), reasoning_error (it found the right code but drew the wrong conclusion), scope_confusion (it answered about the wrong part of the codebase).

Walkthrough

What makes a question diagnostic

A diagnostic question has three characteristics:

  1. It has a verifiable answer. You can check the answer against the actual code. "Is this a well-designed system?" is not verifiable. "Does the Router class inherit from APIRouter?" is verifiable.

  2. It tests one capability at a time. "Find the auth middleware and explain how it interacts with the rate limiter and suggest improvements" tests three things at once. If the system gets it wrong, you can't tell which capability failed. Split it into three questions.

  3. The difficulty is intentional. You should know whether a question is easy (symbol lookup) or hard (change impact) when you write it. If you don't know how hard it is, you can't interpret the results.

10 worked examples

Here are 10 example benchmark questions for a fictional CRUD-style FastAPI application. This is not the real FastAPI framework repo. It's a simplified template app we're using to illustrate the patterns. When you write your own questions, translate the structure (category, difficulty, what it tests) to your actual anchor repo rather than copying these filenames.

Symbol lookup (easy)

Q1

What does the get_current_user function in auth/dependencies.py do?

Gold answer: It's a FastAPI dependency that extracts a JWT token from the Authorization header, decodes it, looks up the user in the database, and returns the User model. If the token is missing or invalid, it raises an HTTPException with status 401.

Design note: Good diagnostic question. One function, one file, verifiable answer. Tests whether the system can find and describe a specific symbol.

Symbol lookup (medium)

Q2

What environment variables does the application require to start?

Gold answer: DATABASE_URL (PostgreSQL connection string), SECRET_KEY (JWT signing key, required, no default), REDIS_URL (optional, defaults to localhost:6379).

Design note: Tests retrieval across multiple files. Env vars are typically scattered across config modules, not in one place.

Architecture explanation

Q3

How does a request flow from the HTTP endpoint to the database for the POST /items route?

Gold answer: The request hits the FastAPI router in routers/items.py, passes through the get_current_user dependency for auth, validates the request body with the ItemCreate Pydantic model, calls crud.items.create_item() which uses SQLAlchemy to insert a row, and returns the created item as an ItemResponse model.

Design note: Tests multi-file, multi-layer understanding. The gold answer traces a specific path through the code.

Architecture explanation

Q4

What is the relationship between models.py, schemas.py, and crud.py in this project?

Gold answer: models.py defines SQLAlchemy ORM models (database tables). schemas.py defines Pydantic models for API request/response validation. crud.py contains database operations that accept Pydantic schemas and return ORM models. The separation keeps database concerns out of the API layer.

Design note: Tests structural understanding, not just symbol lookup.

Change impact

Q5

What would break if I removed the verify_token middleware from main.py?

Gold answer: All authenticated endpoints would stop requiring authentication. Any route that depends on get_current_user would raise an error because the token extraction dependency would no longer find a valid token in the request. The /health and /docs endpoints would be unaffected since they don't use auth dependencies.

Design note: This is harder. It requires reasoning about dependencies and side effects, not just reading code.

Change impact

Q6

If I changed the Item model to add a category field, what other files would I need to update?

Gold answer: You'd need to update schemas.py (add category to ItemCreate and ItemResponse), create a migration (alembic revision --autogenerate), and optionally update crud.py if there are category-specific queries. Tests that reference items would also need updating.

Design note: Tests multi-file impact reasoning.

Debugging

Q7

Users report getting 422 errors when creating items. The request body looks correct. What's likely wrong?

Gold answer: A 422 from FastAPI means Pydantic validation failed. Check whether the ItemCreate schema has required fields the client isn't sending, whether field types mismatch (e.g., sending a string for an integer field), or whether a recent schema change added required fields without updating the client.

Design note: Tests diagnostic reasoning. The gold answer explains the mechanism, not just the fix.

Debugging

Q8

The /search endpoint returns empty results even though there are matching items in the database. Where would you look first?

Gold answer: Check the search query in crud/search.py. Likely a SQL LIKE pattern issue (missing % wildcards), a case sensitivity mismatch, or a filter that's too restrictive. Also check whether the search index is up to date if the project uses full-text search.

Design note: Open-ended enough to test reasoning, specific enough to grade.

Onboarding

Q9

I'm new to this project. Where should I start reading to understand the main API structure?

Gold answer: Start with main.py to see how the app is assembled and which routers are included. Then look at routers/ to see the available endpoints. Read models.py and schemas.py to understand the data shapes. The README.md has setup instructions and the docs/ directory has architecture notes.

Design note: Tests the system's ability to give useful orientation, not just list files.

Onboarding

Q10

How do I run the tests for this project?

Gold answer: Run pytest from the project root. Tests are in the tests/ directory. You need a test database configured. Check conftest.py for the test database URL and fixtures. Run pytest -v for verbose output or pytest tests/test_items.py for a specific module.

Design note: Practical onboarding question. The answer should be actionable, not theoretical.

A bad question, rewritten

Bad: "Is the code well-organized?"

This question is not diagnostic. Any answer could be argued as partially correct. It doesn't test a specific capability and can't be graded consistently.

Better: "What design pattern does the project use to separate database models from API schemas, and where are the boundaries?"

This is verifiable (the pattern either exists or it doesn't), tests architectural understanding, and has a concrete answer you can grade.

Building your grading rubric

For each benchmark question, grade the system's answer using this four-level rubric:

GradeCriteriaAction
Fully correctAll key facts match the gold answer. No fabricated details. Evidence is from the right files.Record as pass.
Partially correctCore idea is right but missing details, or includes minor inaccuracies alongside correct content.Record which parts are correct and which are missing. Apply a failure label to the missing parts.
UnsupportedThe answer sounds plausible but references code, functions, or patterns that don't exist in the repo.Record as unsupported. Apply hallucination or wrong_chunk failure label.
WrongThe answer contradicts what the code actually does.Record as wrong. Apply the most specific failure label.

When you grade, always check against the actual code. Don't grade from memory. The gold answer is your reference, but the code is the ground truth.

Exercises

  1. Review your 10 questions from the previous lesson. For each one, ask: is it diagnostic? Does it test one capability? Can I verify the answer against the code? Rewrite any that fail these tests.

  2. Expand your benchmark set to 30 questions. Use the five categories as a guide. Aim for at least 5 per category. Use the 10 worked examples above as templates.

  3. Write gold answers for at least 8 of your 30 questions. Keep them concise (2-5 sentences) but specific enough to grade against.

  4. Take 3 of your benchmark questions and try answering them by manually prompting a model (using the quickstart from Module 1). Grade the answers using the four-level rubric. Apply failure labels to any non-fully-correct answers.

  5. Save your benchmark set as a structured file:

# Create in your anchor repo directory
touch benchmark-questions.jsonl
{"id": "q01", "category": "symbol_lookup", "question": "What does get_current_user do?", "gold_answer": "It extracts a JWT...", "difficulty": "easy"}
{"id": "q02", "category": "architecture", "question": "How does a request flow from...", "gold_answer": "The request hits...", "difficulty": "medium"}

Completion checkpoint

You should now have:

  • 30+ benchmark questions covering all five categories
  • Gold answers for at least 8 questions
  • At least 3 questions hand-graded using the four-level rubric with failure labels
  • A benchmark-questions.jsonl file with your structured benchmark set
  • Confidence that your questions are diagnostic, verifiable, and categorized by difficulty
  • "Golden" by Huntrix queued in your playlist

What's next

Run Logs and Baseline. Once the benchmark is solid, the next lesson gives you the schema and process for recording runs, grading them, and establishing the number you will spend the rest of the path improving.

References

Start here

Build with this

Deep dive

  • LMSYS: Chatbot Arena — large-scale model evaluation; useful for understanding how evaluation methodology affects conclusions
Your Notes
GitHub Sync

Sync your lesson notes to a private GitHub Gist. If you have not entered a token yet, the sync button will open the GitHub token modal.
Glossary
API (Application Programming Interface)Foundational terms
A structured way for programs to communicate. In this context, usually an HTTP endpoint you call to interact with an LLM.
AST (Abstract Syntax Tree)Foundational terms
A tree representation of source code structure. Used by parsers like Tree-sitter to understand code as a hierarchy of functions, classes, and statements. You'll encounter this more deeply in the Code Retrieval module, but the concept appears briefly in retrieval fundamentals.
BM25 (Best Match 25)Foundational terms
A classical ranking function for keyword search. Scores documents by term frequency and inverse document frequency. Often competitive with or complementary to vector search.
ChunkingFoundational terms
Splitting a document into smaller pieces for indexing and retrieval. Chunk boundaries significantly affect retrieval quality. Split at the wrong place and your retrieval will return half a function or the end of one paragraph glued to the start of another.
Context engineeringFoundational terms
The discipline of selecting, packaging, and budgeting the information a model sees at inference time. Prompts, retrieved evidence, tool results, memory, and state are all parts of context. Context engineering is arguably the core skill of AI engineering. Bigger context windows are not a substitute for better context selection.
Context rotFoundational terms
Degradation of output quality caused by stale, noisy, or accumulated context. Symptoms include stale memory facts, conflicting retrieved evidence, bloated prompt history, and accumulated instructions that contradict each other. A form of technical debt in AI systems.
Context windowFoundational terms
The maximum number of tokens an LLM can process in a single request (input + output combined).
EmbeddingFoundational terms
A fixed-length numeric vector representing a piece of text. Used for similarity search: texts with similar meanings have nearby embeddings.
EndpointFoundational terms
A specific URL path that accepts requests and returns responses (e.g., POST /v1/chat/completions).
GGUFFoundational terms
A file format for quantized models used by llama.cpp and Ollama. When you see a model name like qwen2.5:7b-q4_K_M, the suffix indicates the quantization scheme. GGUF supports mixed quantization (different precision for different layers) and is the most common format for local inference.
HallucinationFoundational terms
When a model generates content that sounds confident but isn't supported by the evidence it was given, or fabricates details that don't exist. Not the same as "any wrong answer"; a model that misinterprets ambiguous instructions gave a bad answer but didn't hallucinate. Common causes: weak prompt, missing context, context rot, model limitation, or retrieval failure.
InferenceFoundational terms
Running a trained model to generate output from input. What happens when you call an API. Most AI engineering work is inference-time work: building systems around models, not training them. Use "inference," not "inferencing."
JSON (JavaScript Object Notation)Foundational terms
A lightweight text format for structured data. The lingua franca of API communication.
Lexical searchFoundational terms
Finding items by matching keywords or terms. Includes BM25, TF-IDF (Term Frequency–Inverse Document Frequency), and simple keyword matching. Returns exact term matches, not semantic similarity.
LLM (Large Language Model)Foundational terms
A neural network trained on large text corpora that generates text by predicting the next token. The core technology behind AI engineering; every tool, pattern, and pipeline in this curriculum runs on top of one.
MetadataFoundational terms
Structured information about a document or chunk (file path, language, author, date, symbol type). Used for filtering retrieval results.
Neural networkFoundational terms
A computing system loosely inspired by biological neurons, built from layers of mathematical functions that transform inputs into outputs. LLMs are a specific type of neural network (transformers) trained on text. You don't need to understand neural network internals to do AI engineering, but knowing the term helps when reading external resources.
Reasoning modelFoundational terms
A model optimized for complex multi-step planning, math, and logic (e.g., o3, o4-mini). Slower and more expensive but better on hard problems. Sometimes called "LRM" (large reasoning model), but "reasoning model" is the more consistent term across provider docs.
RerankingFoundational terms
A second-pass scoring step that re-orders retrieved results using a more expensive model. Improves precision after an initial broad retrieval.
SchemaFoundational terms
A formal description of the shape and types of a data structure. Used to validate inputs and outputs.
SLM (small language model)Foundational terms
A compact model (typically 1-7B parameters) that runs on consumer hardware with lower cost, latency, and better privacy (e.g., Phi, small Llama variants, Gemma). The right choice when privacy, offline operation, predictable cost, or low latency matter more than peak capability.
System promptFoundational terms
A special message that sets the model's behavior, role, and constraints for a conversation.
TemperatureFoundational terms
A parameter controlling output randomness. Lower values produce more deterministic output; higher values produce more varied output. Does not affect the model's intelligence.
TokenFoundational terms
The basic unit an LLM processes. Not a word. Tokens are sub-word fragments. "unhappiness" might be three tokens: "un", "happi", "ness". Token count determines cost and context window usage.
Top-kFoundational terms
The number of results returned from a retrieval query. "Top-5" means the five highest-scoring results.
Top-p (nucleus sampling)Foundational terms
An alternative to temperature for controlling output diversity. Selects from the smallest set of tokens whose cumulative probability exceeds p.
Vector searchFoundational terms
Finding items by proximity in embedding space (nearest neighbors). Returns "similar" results, not "exact match" results.
vLLM (virtual LLM)Foundational terms
An inference serving engine (not a model) that hosts open-weight models behind an OpenAI-compatible HTTP endpoint. Infrastructure layer, not model layer. Relevant when moving from hosted APIs to self-hosting.
WeightsFoundational terms
The learned parameters inside a model. Changed during training, fixed during inference.
Workhorse modelFoundational terms
A general-purpose LLM optimized for speed and broad capability (e.g., GPT-4o-mini, Claude Haiku, Gemini Flash). The default for most tasks. When someone says "LLM" without qualification, they usually mean this.
BaselineBenchmark and Harness terms
The first measured performance of your system on a benchmark. Everything else is compared against this. Without a baseline, you can't tell whether a change helped.
BenchmarkBenchmark and Harness terms
A fixed set of questions or tasks with known-good answers, used to measure system performance over time.
Run logBenchmark and Harness terms
A structured record (typically JSONL) of every system run: what input was given, what output was produced, what tools were called, how long it took, and what it cost. The raw data that evals, telemetry, and cost analysis are built from.
A2A (Agent-to-Agent protocol)Agent and Tool Building terms
An open protocol for peer-to-peer agent collaboration. Agents discover each other's capabilities and delegate or negotiate tasks as equals. Different from MCP (which connects agents to tools, not to other agents) and from handoffs (which transfer control within one system).
AgentAgent and Tool Building terms
A system where an LLM decides which tools to call, observes results, and iterates until a task is complete. Agent = model + tools + control loop.
Control loopAgent and Tool Building terms
The code that manages the agent's cycle: send prompt, check for tool calls, execute tools, append results, repeat or finish.
HandoffAgent and Tool Building terms
Passing control from one agent or specialist to another within an orchestrated system.
MCP (Model Context Protocol)Agent and Tool Building terms
An open protocol for exposing tools, resources, and prompts to AI applications in a standardized way. Connects agents to capabilities (tools and data), not to other agents.
Tool calling / function callingAgent and Tool Building terms
The model's ability to request execution of a specific function with structured arguments, rather than just generating text.
Context compilation / context packingCode Retrieval terms
The process of selecting and assembling the smallest useful set of evidence for a specific task. Not "dump everything retrieved into the prompt."
GroundingCode Retrieval terms
Tying model assertions to specific evidence. A grounded answer cites what it found; an ungrounded answer asserts without evidence.
Hybrid retrievalCode Retrieval terms
Combining multiple retrieval methods (e.g., vector search + keyword search + metadata filters) and merging or reranking the results.
Knowledge graphCode Retrieval terms
A data structure that stores entities and their relationships explicitly (e.g., "function A calls function B," "module X imports module Y"). Useful for traversal and dependency reasoning. One retrieval strategy among several, often overused when simpler metadata or adjacency tables would suffice.
RAG (Retrieval-Augmented Generation)Code Retrieval terms
A pattern where the model's response is grounded in retrieved external evidence rather than relying solely on its training data.
Symbol tableCode Retrieval terms
A mapping of code identifiers (functions, classes, variables) to their locations and metadata.
Tree-sitterCode Retrieval terms
An incremental parsing library that builds ASTs for source code. Used in this curriculum for code-aware chunking and symbol extraction.
Context packRAG and Grounded Answers terms
A structured bundle of evidence assembled for a specific task, with metadata about provenance, relevance, and token budget.
Evidence bundleRAG and Grounded Answers terms
A collection of retrieved items grouped for a specific sub-task, with enough metadata to evaluate whether the evidence is relevant and sufficient.
Retrieval routingRAG and Grounded Answers terms
Deciding which retrieval strategy or method to use for a given query. Different questions need different retrieval methods.
EvalObservability and Evals terms
A structured test that measures system quality. Not the same as training. Evals measure, they don't change the model.
Harness (AI harness / eval harness)Observability and Evals terms
The experiment and evaluation framework around your model or agent. It runs benchmark tasks, captures outputs, logs traces, grades results, and compares system versions. It turns ad hoc "try it and see" into repeatable, comparable experiments. Typically includes: input dataset, prompt and tool configuration, model/provider selection, execution loop, logging, grading, and artifact capture.
LLM-as-judgeObservability and Evals terms
Using a language model to evaluate or grade the output of another model or system. Useful for scaling evaluation beyond manual review, but requires rubric quality, judge consistency checks, and human spot-checking. Not a replacement for exact-match checks where they apply.
OpenTelemetry (OTel)Observability and Evals terms
An open standard for collecting and exporting telemetry data (traces, metrics, logs). Vendor-agnostic.
RAGASObservability and Evals terms
A specific eval framework for retrieval-augmented generation. Measures metrics like faithfulness, relevance, and context precision. One tool example, not a foundational concept. Learn the metrics first, then the tool.
SpanObservability and Evals terms
A single operation within a trace (e.g., one tool call, one retrieval query). Traces are made of spans.
TelemetryObservability and Evals terms
Structured data about system behavior: what happened, when, how long it took, what it cost. Includes traces, metrics, and events.
TraceObservability and Evals terms
A structured record of one complete run through the system, including all steps, tool calls, and decisions.
Long-term memoryOrchestration and Memory terms
Persistent facts that survive across conversations. Requires write policies to manage what gets stored, updated, or deleted.
OrchestrationOrchestration and Memory terms
Explicit control over how tasks are routed, delegated, and synthesized across multiple agents or specialists.
RouterOrchestration and Memory terms
A component that decides which specialist or workflow path to use for a given query.
SpecialistOrchestration and Memory terms
An agent or workflow tuned for a narrow task (e.g., "code search," "documentation lookup," "test generation"). Specialists are composed by an orchestrator.
Thread memoryOrchestration and Memory terms
Conversation state that persists within a single session or thread.
Workflow memoryOrchestration and Memory terms
Intermediate state that persists within a multi-step task but doesn't survive beyond the workflow's completion.
Catastrophic forgettingOptimization terms
When fine-tuning causes a model to lose capabilities it had before training. The model gets better at the fine-tuned task but worse at tasks it previously handled. PEFT methods like LoRA reduce this risk by freezing original weights.
DistillationOptimization terms
Training a smaller (student) model to reproduce the behavior of a larger (teacher) model on a specific task.
DPO (Direct Preference Optimization)Optimization terms
A method for preference-based model optimization that's simpler than RLHF, training the model directly on preference pairs without a separate reward model.
Fine-tuningOptimization terms
Updating a model's weights on task-specific data to change its behavior permanently. An umbrella term that includes SFT, instruction tuning, RLHF, DPO, and other techniques. See the fine-tuning landscape table in Lesson 8.3 for how these relate.
Full fine-tuningOptimization terms
Updating all of a model's parameters during training, as opposed to PEFT methods that update only a small subset. Requires significantly more GPU memory and compute. Produces the most thorough adaptation but carries higher risk of catastrophic forgetting.
Inference serverOptimization terms
Software (like vLLM or Ollama) that hosts a model and serves inference requests.
Instruction tuningOptimization terms
A specific application of SFT where the training data consists of instruction-response pairs. This is how base models become chat models: the technique is SFT, the data format is instructions. Not a separate technique from SFT.
LoRA (Low-Rank Adaptation)Optimization terms
A parameter-efficient fine-tuning method that trains small adapter matrices instead of updating all model weights. Dramatically reduces GPU memory and compute requirements.
Parameter countOptimization terms
The number of learned weights in a model, commonly expressed in billions (e.g., "7B" = 7 billion parameters). Determines memory requirements (roughly 2 bytes per parameter at FP16) and broadly correlates with capability, though training quality and architecture matter as much as size. See Model Selection and Serving for sizing guidance.
PEFT (Parameter-Efficient Fine-Tuning)Optimization terms
A family of methods (including LoRA) that fine-tune a small subset of parameters instead of the full model.
Preference optimizationOptimization terms
Training methods (RLHF, DPO) that use human or automated preference signals to improve model behavior. "This output is better than that output" rather than "this is the correct output."
QLoRA (Quantized LoRA)Optimization terms
LoRA applied to a quantized (compressed) base model. Further reduces memory requirements, enabling fine-tuning on consumer hardware.
QuantizationOptimization terms
Reducing the precision of model weights (e.g., FP16 → INT4) to shrink memory usage and increase inference speed at some quality cost. A 7B model at FP16 needs ~14 GB VRAM; quantized to 4-bit, it fits in ~4 GB. Common formats include GGUF (llama.cpp/Ollama), GPTQ and AWQ (vLLM/HuggingFace). See Model Selection and Serving for format details and tradeoffs.
OverfittingOptimization terms
When a model memorizes training examples instead of learning generalizable patterns. The model performs well on training data but poorly on new inputs. Detected by monitoring validation loss alongside training loss.
RLHF (Reinforcement Learning from Human Feedback)Optimization terms
A training method that uses human preference signals to improve model behavior through a reward model. More complex than DPO (requires training a separate reward model) but offers more control over the optimization objective.
SFT (Supervised Fine-Tuning)Optimization terms
Fine-tuning using input-output pairs where the desired output is known. The most common fine-tuning approach.
TRL (Transformer Reinforcement Learning)Optimization terms
A Hugging Face library for training language models with reinforcement learning, SFT, and other optimization methods.
Consumer chat appCross-cutting terms
The browser or desktop product meant for human conversation (ChatGPT, Claude, HuggingChat). Useful for experimentation, but not the same as API access.
Developer platformCross-cutting terms
The provider's API, billing, API-key, and developer-docs surface. This is what you need for this learning path.
Hosted APICross-cutting terms
The provider runs the model for you and you call it over HTTP.
Local inferenceCross-cutting terms
You run the model on your own machine.
ProviderCross-cutting terms
The company or service that hosts a model API you call from code.
Prompt cachingCross-cutting terms
Reusing computation from repeated prompt prefixes to reduce latency and cost on subsequent requests with the same prefix.
Rate limitingCross-cutting terms
Constraints on how many API requests you can make per unit of time. An operational concern that affects system design and cost.
Token budgetCross-cutting terms
The maximum number of tokens you allocate for a specific part of the context (e.g., "retrieval evidence gets at most 4K tokens"). A context engineering tool for preventing any single component from dominating the context window.