# What is ground truth in LLM evaluation?

> Ground truth is the trusted reference answer or label an LLM output is scored against. What it looks like, how to build it and what to do without it.

Source: https://www.honeyhive.ai/learn/ground-truth-llm-evaluation (updated 2026-11-12)

Ground truth in LLM evaluation is the reference answer or label you trust as correct and score a model's output against. AWS authors [define](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) ground truth data as data "known to be factual," and say it makes deterministic evaluation and custom benchmarks possible.

It sits on one example, not on the whole set. A set of inputs paired with reviewed references is a [golden dataset](/learn/golden-dataset). Ground truth is the reference half of each pair.

## What ground truth looks like

The form depends on the task and on how you score.

| Task | Reference | How it is scored |
|---|---|---|
| Classification, such as sentiment | A label. Anthropic's [example](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests) uses 1,000 tweets with human-labeled sentiments. | Exact match, which Anthropic [defines](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests) as checking the output against a predefined correct answer after normalizing whitespace and case. |
| Question answering | An expert-written answer. OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) lists "hard-coded correct answers to questions created by domain experts". | Exact match, or an LLM grader given the answer. |
| Summarization | A reference summary. Anthropic's [example](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests) uses 200 articles with reference summaries. | Overlap metrics such as ROUGE-L, or an LLM grader. |
| Open-ended generation | A "gold standard" answer. OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) describes reference-guided grading, where the judge uses it "as a benchmark". | An LLM judge compares the response to the reference. |

The AWS authors add that the metric decides the shape of the reference: they [write](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) that "the evaluation metric's implementation must inform ground truth curation." Their format is question, answer and fact triplets, with facts of three words or fewer and alternative forms joined by `<OR>`, as in "1,234 million<OR>1.234 billion". See [LLM evaluation metrics](/learn/llm-evaluation-metrics) for how metrics differ.

## Why it is hard to get

Three problems come up again and again.

- **Cost.** OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) says expert human annotators providing ground-truth labels is "expensive and time-consuming." The AWS authors [say](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) expert-only generation across a large knowledge base becomes "prohibitively time and resource intensive."
- **Disagreement.** Experts do not always agree. OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) suggests multiple rounds of review to refine the scorecard and consensus votes to aggregate reviewers. Anthropic's [advice](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests) is to include ambiguous cases where even humans struggle to agree. Those are the cases where a single reference is weakest. Measure it with [inter-annotator agreement](/learn/inter-annotator-agreement).
- **Drift.** Microsoft's [guidance](https://github.com/microsoft/promptflow-resource-hub/blob/main/sample_gallery/golden_dataset/copilot-golden-dataset-creation-guidance.md) says that when quality keeps shifting as grounding data, model or prompt changes, the reference set "is no longer up to date with reality and should be revised."

## How to build it

1. **Start small and expert-made.** The AWS authors [recommend](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) a small, high-signal set curated by subject matter experts, which also aligns stakeholders on the questions that matter.
2. **Cover real use.** Anthropic [says](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests) to design evals that "mirror your real-world task distribution" and to include edge cases.
3. **Make the answer complete.** Microsoft's [guidance](https://github.com/microsoft/promptflow-resource-hub/blob/main/sample_gallery/golden_dataset/copilot-golden-dataset-creation-guidance.md) wants the expert answer to be "both correct and complete."
4. **Let an LLM draft, not decide.** Microsoft's [guidance](https://github.com/microsoft/promptflow-resource-hub/blob/main/sample_gallery/golden_dataset/copilot-golden-dataset-creation-guidance.md) has experts review and correct any LLM-drafted answer, with an optional second expert pass. See [synthetic test data for evals](/learn/synthetic-test-data-for-evals) for generating inputs.
5. **Scale by risk.** The AWS authors [pair](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) a risk-based approach with LLM prompting to grow the set past what experts can write.

## When there is no ground truth

Open-ended outputs and live traffic often have no single right answer. You can still score them without a reference. OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) describes single-answer grading, which assigns "a score or rating based on predefined quality metrics", and pairwise comparison, which picks the better of two responses. Both are [LLM-as-a-judge](/learn/llm-as-a-judge) setups. The same guide says to scale up only once the judge "consistently agrees with human annotations."

Zheng and colleagues [reported](https://arxiv.org/abs/2306.05685) that GPT-4 as a judge reached over 80% agreement with humans, the same level humans reach with each other, and also listed limited reasoning ability among the limits of LLM judges. A small set of human-labeled ground truth is how you check a judge. See [LLM judge vs human evaluation](/learn/llm-judge-vs-human-evaluation) and [online vs offline evaluation](/learn/online-vs-offline-evaluation).

## Keeping ground truth in HoneyHive

HoneyHive's [dataset docs](https://docs.honeyhive.ai/v2/datasets/introduction.md) split each datapoint into input, ground truth, chat history and metadata. Ground truth fields "are used by evaluators to compare against the actual output of your application", and metadata is not used as input or ground truth. You can [upload](https://docs.honeyhive.ai/v2/datasets/import.md) JSON, JSONL or CSV files that include ground truth, or [curate](https://docs.honeyhive.ai/v2/datasets/dataset-curation.md) datapoints from production traces, which keep a `linked_event` reference back to the original trace. [Annotation queues](https://docs.honeyhive.ai/v2/evaluation/annotation-queues.md) collect expert labels and turn judgments into ground truth for evaluators.

## FAQ

### What is ground truth in LLM evaluation?

The reference answer or label you treat as correct and score model outputs against. HoneyHive's [dataset docs](https://docs.honeyhive.ai/v2/datasets/introduction.md) describe it as "the expected or ideal outputs or reference answers for a given input."

### Is ground truth the same as a golden dataset?

No. Ground truth is the reference value on one example. A golden dataset is a reviewed collection of inputs with those references. See [golden dataset](/learn/golden-dataset).

### Can an LLM generate ground truth?

It can draft it, but a person must check it. AWS authors [wrote](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) that "LLM-generated ground truth isn't a substitute for use case SME involvement."

### What if I have no ground truth?

Score without a reference. OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) describes single-answer grading against predefined quality metrics and pairwise comparison, and advises using human feedback to calibrate automated scoring. See [LLM as a judge](/learn/llm-as-a-judge).

### Why is ground truth hard to create?

Experts are slow and costly, and they sometimes disagree. OpenAI's [guide](https://developers.openai.com/api/docs/guides/evaluation-best-practices) calls expert annotation "expensive and time-consuming", and AWS authors [say](https://aws.amazon.com/blogs/machine-learning/ground-truth-generation-and-review-best-practices-for-evaluating-generative-ai-question-answering-with-fmeval/) expert-only generation becomes "prohibitively time and resource intensive" at enterprise scale.