Figure 11 Prompt used to train a rewriter model. We prepend the suffix with the last five words from the prefix to ensure the rewriter always smoothly continues the prefix.
Rewriter prompt

Provided is a Context C and a continuation of the context X.

Your task is to provide a REWRITTEN continuation that is a valid continuation of the context C and superior to the continuation X. If the continuation X contains factual inaccuracies or hallucinations, they should be corrected. If the continuation X contains an unsafe response, this should be corrected to instead reply as safely as possible, while being as helpful/useful as possible. If the continuation X is already a good continuation, you should simply copy it as the REWRITTEN continuation.

Provide only a rewrite that matches the length of the continuation X (128 tokens), without any additional text or explanations.

<Continuation X start>{prefix_ending}{suffix}<Continuation X end>
<Context C start>{prefix}<Context C end>
<Rewritten continuation start>{prefix_ending}

A.2.1 Factuality Prompts: with reference

Base Factuality Prompt - with reference

You are a factuality evaluator.

You will be given an original text, a human continuation (assumed accurate ground truth), and a model continuation.

Your task is to determine if the model continuation contains any hallucinations, internal inconsistencies, or statements implausible given the original text and the human continuation.

Instructions:
  1. Reason step by step about whether the model continuation logically follows from the original text and the human continuation.
  2. Note any hallucinations, internal contradictions, or statements that are implausible.
  3. At the end, output your reasoning and final label strictly in JSON format.
JSON Output Format:
{ "reasoning": "your reasoning here", "label": "No Hallucination" | "Possible Hallucination" | "Definite Hallucination" }

Prompt Variants v1-v5 We now describe for each prompt version we tried how it differs from the base prompt provided above.

V1 Differences from Base:

  • Explicitly includes the human continuation as ground truth and instructs the evaluator to treat it as the primary reference.
  • Emphasizes step-by-step reasoning about whether the model continuation logically follows from both the original text and the human continuation.
  • Focuses on hallucinations, internal inconsistencies, or statements implausible given both references.

V2 Differences from Base:

21