🎯 场景：把 1 条种子指令改写成 N 个变体——同任务、不同表面（措辞、register、长度提示）。SFT 数据从手工种子规模化扩展的纵向维度。

Quick Use

Use when: You want to rewrite ONE instruction into N variants that preserve the underlying task but vary surface form, register, or style. Fill in: {{seed_instruction}} = the original instruction; {{n_variants}} = how many rewrites to produce. You'll get: A list of variants, each with a delta describing what changed compared to the seed. Output is JSON.

Purpose

Take one seed instruction from a small high-quality SFT set and produce N diverse rewrites that preserve task semantics but vary surface form, register, and elicitation style. Used to multiply hand-curated SFT seeds into a larger training set without drifting off-task. Output is structured so each variant can be paired with the original gold answer (if the answer is invariant to rewrites) or sent to a separate answer-generation pass.

Prompt

You expand a single seed instruction into {{n_variants}} diverse rewrites.

Constraints:
- Each variant MUST preserve the same underlying task and produce the same
  correct answer as the seed (no scope drift, no added or removed fields).
- Variants should differ on at least two of: surface phrasing, register
  (formal/casual/terse), context (with/without backstory), output format hint,
  language complexity.
- Do NOT introduce new entities, numbers, constraints, or topics not present
  in the seed.

Seed instruction:
{{seed_instruction}}

Return ONLY this JSON object:
{
  "variants": [
    {"text": "<variant 1>", "delta": "<one short phrase: what changed vs seed>"},
    {"text": "<variant 2>", "delta": "<...>"}
    /* ... up to {{n_variants}} entries */
  ]
}

Example

Input:

seed_instruction: "Summarize the following article in 3 bullet points."
n_variants: 3

Expected output:

{
  "variants": [
    {"text": "Give me 3 bullets that capture the key points of this article.", "delta": "casual register"},
    {"text": "Please produce a three-bullet summary of the article below, preserving the original ordering of major points.", "delta": "formal + ordering hint"},
    {"text": "TL;DR in 3 bullets:", "delta": "terse style with format cue"}
  ]
}

Failure Modes

• Scope drift — variant changes the task ("Summarize in 3 bullets" → "Translate the article"). Mitigation: post-filter by re-asking a verifier model "is variant X the same task as seed Y?" and dropping disagreements.
• Cosmetic-only diversity — all variants are the same sentence with punctuation tweaks. Detect with cheap surface metrics (Jaccard on tokens); reject batches with mean similarity > 0.85.
• New constraint injection — variant adds "in less than 50 words" or "in JSON" when seed had no such constraint. The delta field exposes this for review.
• Answer divergence — when paired with the seed's gold answer, the variant actually produces a different correct answer. Mitigation: re-run the gold answer generator on each variant and compare.

Tuning Notes

• 模型差异：frontier 模型多样性显著高于开源中等模型；如果用 7B 模型生成会 得到大量"换标点"伪多样性，建议用 GPT-4/Claude Sonnet 类模型生成 seed→variant 扩展。
• 温度：0.8–1.0，多样性优先；之后用低温 verifier 过滤。
• N 选择：每个 seed 扩 5–10 个，过滤后保留率约 40–70%，得到 2–7 个有效 variant。
• 与人工的关系：此卡是规模化工具，不替代人工审核。建议保留 5–10% 样本送人工 采样审核。

Changelog

• 0.1.0 — Initial card.