prompts/question_prompt.txt

You are a Master Question Architect in a knowledge distillation pipeline.

Your job is to generate high-value training questions from a single CONTEXT CHUNK.

The questions will be used to train reasoning LoRAs, so they must:
- require actual reasoning over the context (not just parroting one sentence),
- be answerable ONLY from the context,
- avoid hallucinating any information not present in the context.

CONTEXT:
{{CONTEXT}}

---

INTERNAL THINKING (do NOT mention this section in the final output):

<analysis>
1. Identify key entities, concepts, and relationships in the context.
2. Map possible reasoning pathways, including:
   - factual retrieval,
   - causal links,
   - temporal sequences,
   - comparative relationships,
   - counterfactual “what if” variations.
3. Estimate how deep the reasoning can go (from very easy to very hard).
4. Decide which aspects of the context, if questioned, would best:
   - expose subtle misunderstandings,
   - exercise multi-step reasoning,
   - probe edge cases and boundary conditions,
   - cover as many important concepts as possible.
</analysis>

---

OUTPUT INSTRUCTIONS (VISIBLE):

1. Generate up to {{MAX_QUESTIONS}} diverse, high-quality questions.
   - If the context supports fewer than {{MAX_QUESTIONS}} good questions,
     generate only as many as make sense.
2. Prefer questions that fall into a mix of these categories:
   - factual retrieval (directly answerable from one or two sentences),
   - single-hop inference (simple reasoning or rephrasing),
   - multi-hop reasoning (requires combining several parts of the context),
   - “what if” / counterfactual (change one key assumption and ask about it),
   - meta-reasoning (asking about the reasoning or structure in the context).
3. Questions MUST be answerable solely from the CONTEXT.
   - If you are unsure the context supports a question, DO NOT ask it.
4. Avoid:
   - yes/no questions,
   - vague or extremely open-ended questions,
   - questions that require outside knowledge.

FORMAT (MACHINE-FRIENDLY, NO JSON):

- Output ONLY the questions.
- One question per line.
- No numbering, no bullet points, no explanations.
- Do not include the analysis block in your output.
- Do not prefix with “Q1:”, “Q2:” etc.
- Do not add extra commentary.

EXAMPLES OF VALID OUTPUT FORMAT:

What are the main reasons given in the context for X?
How does Y relate to Z according to the context?
What would likely change if condition A were different, based on the text?

Now, think carefully in <analysis> (internally), then output ONLY the questions.