Faithfulness Score Calibration Prompt

This prompt is for evaluation leads and AI engineers who need to align an LLM judge's faithfulness scores with human ratings. The core job is calibration: you have a set of human-annotated examples where each AI output has a known groundedness score, and you need the LLM judge to produce scores that match the human distribution. Without calibration, LLM judges drift—scoring too leniently on some model versions, too harshly on others, or shifting their interpretation of the rubric over multiple evaluation runs. This prompt solves that by presenting few-shot calibration examples with human-annotated ground truth before the judge scores new items.

Use this prompt when you have at least 20–50 human-annotated calibration examples covering the full range of your groundedness rubric (e.g., scores 1–5). The calibration examples should include edge cases: partially supported claims, ambiguous evidence relationships, and outputs where the model correctly abstained. Do not use this prompt when you lack human annotations, when your rubric is still in draft, or when you need real-time scoring latency below 500ms—calibration adds token overhead. This prompt is also inappropriate for binary pass/fail gates; use the Pass/Fail Groundedness Gate Prompt instead.

The prompt works best as a pre-scoring step in batch evaluation pipelines. Run it once per evaluation session to establish judge alignment, then feed the calibrated judge your unlabeled outputs. Store the calibration examples in your eval dataset with versioned human annotations so you can detect when the judge's alignment drifts over time. If you change your groundedness rubric, you must update the calibration examples and re-run alignment—old calibration data will silently corrupt new scores. Always log the calibration examples used, the judge's raw scores, and the human reference scores for auditability.

This template is the core of the calibration workflow. It provides a structured prompt that you can copy, adapt, and embed directly into your evaluation harness. The prompt instructs an LLM judge to score the faithfulness of a generated answer against a source document, but it does so by first presenting a set of few-shot examples with human-annotated scores. These examples act as the calibration anchor, teaching the judge what a '1', '3', or '5' means in your specific context before it scores the target item. The square-bracket placeholders are designed to be filled by your application code at runtime, pulling from a calibration dataset that you maintain and version alongside your prompt.

textCopy
You are a strict faithfulness evaluator. Your task is to score how factually consistent a [GENERATED_ANSWER] is with a [SOURCE_DOCUMENT]. A score of 1 means the answer is completely unsupported or contradicts the source. A score of 5 means the answer is fully supported with no added or distorted information.

Before scoring the target, study these calibration examples. Each example includes a source, an answer, a human-annotated score, and a brief justification. Use these to calibrate your internal scoring scale.

[CALIBRATION_EXAMPLES]

---

Now, evaluate the following target item. Apply the same scoring standards you observed in the calibration examples.

[SOURCE_DOCUMENT]:
[SOURCE_TEXT]

[GENERATED_ANSWER]:
[ANSWER_TEXT]

[OUTPUT_SCHEMA]:
Return your evaluation as a single JSON object with the following keys:
- "faithfulness_score": An integer from 1 to 5.
- "justification": A string explaining the score, referencing specific claims and source evidence.
- "uncertainty_flags": An array of strings noting any ambiguous cases or borderline decisions.

[CONSTRAINTS]:
- Do not critique style, grammar, or helpfulness. Only evaluate factual consistency with the source.
- If the answer contains multiple claims, each unsupported claim should lower the score.
- A score of 3 is for answers that are partially supported but contain minor unsupported details or omissions.

To adapt this template, start by curating your [CALIBRATION_EXAMPLES]. This is the most critical variable. Each example should be a clear, concise block of text containing a source snippet, an answer, the human-annotated score, and a short justification. Aim for 3-5 examples that span the full range of your scoring rubric, especially the boundaries between scores (e.g., a clear '2' vs. a clear '3'). Store these examples in a version-controlled file or database so you can update them when you refine your rubric or discover judge drift. The [OUTPUT_SCHEMA] enforces a structured JSON response, which is essential for automated parsing in a CI/CD pipeline. If your evaluation harness uses a different schema, replace it, but always include a justification field for debugging. The [CONSTRAINTS] section is your defense against scope creep; it explicitly tells the judge to ignore style and focus solely on factual consistency, preventing conflation with quality or helpfulness metrics. Before deploying, run this prompt against a held-out set of human-annotated examples to measure the judge's alignment and set an acceptable error tolerance.

The Faithfulness Score Calibration Prompt is designed to be embedded in an automated evaluation harness, not run as a one-off chat. The prompt takes a set of few-shot examples with human-annotated ground truth scores and uses them to calibrate the LLM judge before scoring new outputs. The harness must manage example selection, score extraction, validation, and alignment drift detection across evaluation runs. Treat this prompt as a calibration step that runs before each batch evaluation, not as a standalone scorer—the calibration examples establish the scoring contract for that session.

To wire this into an application, build a calibration runner that executes the prompt with a curated set of 5–15 human-annotated examples covering the full score range. Extract the numeric score from the model response using a strict regex or JSON parser—expect formats like Score: 4 or structured JSON with a faithfulness_score field. Validate that the extracted score falls within the defined scale (e.g., 1–5) and log any parse failures for review. Run the calibration prompt at the start of each evaluation batch and compare the judge's scores on held-out calibration examples against human ground truth. If the mean absolute error exceeds your threshold (commonly 0.5 on a 5-point scale), flag the judge for recalibration or human review before trusting batch results. Store calibration metadata—example IDs, human scores, judge scores, and drift metrics—alongside evaluation results for auditability.

Model choice matters here. Use a model with strong instruction-following and consistent scoring behavior, such as GPT-4, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Avoid smaller or cheaper models for calibration unless you've validated their score stability across runs. Set temperature to 0 or near-zero to reduce variance. Implement a retry policy: if the model returns an unparseable score or refuses to score, retry once with the same prompt. If the second attempt fails, log the failure, exclude that item from batch results, and alert the evaluation pipeline owner. For high-stakes evaluation pipelines—such as those gating model releases or measuring production quality—require human review of calibration drift alerts and maintain a rotating set of calibration examples to prevent overfitting to a fixed few-shot set.

The calibration prompt should be versioned alongside your evaluation code. When you update the prompt, re-baseline against your human-annotated calibration set and document any score distribution shifts. If calibration scores drift after a model provider update, run a pairwise comparison between the old and new judge outputs on the same calibration set to diagnose whether the drift is systematic (requiring recalibration) or random (requiring a model rollback). Avoid running calibration once and caching the results indefinitely—recalibrate at the start of each evaluation run or at minimum daily for continuous monitoring pipelines. The next step after calibration is to feed the calibrated scoring instructions into your batch faithfulness evaluation prompt, using the same scale and criteria the judge just aligned to.