🎯 场景：Self-Instruct 技术——从 5-10 条种子指令出发生成同 task family 的新指令（不是改写）。SFT 数据从手工种子规模化的横向维度，和 variant-expander 互补。

Quick Use

Use when: You have a small bank of seed instructions and want to generate NEW instructions in the same task family (Self-Instruct technique). Fill in: {{seed_instructions}} = JSON array of 5-10 seed examples; {{task_family_description}} = optional one-line description; {{n_new_instructions}} = how many new ones to generate. You'll get: A list of new instructions each with task-family fit, diversity axes, and a novelty check. Output is JSON.

Purpose

Generate new, diverse instructions in the same task family as a small hand-curated seed bank — the Self-Instruct technique. Used to scale a high-quality seed set (50-200 instructions) into a large SFT instruction pool (thousands), without the variants being mere rewrites of the seeds. Distinct from sft/instruction-variant-expander: that card rephrases one seed; this card invents new instructions in the seeds' task family. Output is structured so each new instruction can be filtered for quality before paired with a response.

Prompt

You generate new instructions in the same task family as a set of seed
examples. Goal: produce diverse, novel instructions that a downstream
SFT pipeline can pair with a response.

Seed instructions:
{{seed_instructions}}

Task family description (may be empty; if so, infer from seeds):
{{task_family_description}}

Generate {{n_new_instructions}} new instructions.

Rules:
1. Each new instruction must belong to the same task family as the seeds
   (the same kind of work, not a different task type).
2. Each new instruction must be NEW — not a rewrite of any seed and not
   a near-duplicate of another newly generated instruction.
3. Vary along at least two of: domain / topic, level of specificity,
   register (formal/casual/terse), length of expected response,
   complexity (single-step vs multi-step).
4. Each instruction must be self-contained and answerable on its own
   without external context that the seeds did not require.
5. Do NOT generate instructions that violate docs/SAFETY.md (no harm
   enabling, no PII, no jailbreak prompts).
6. If the task family is too narrow to support {{n_new_instructions}}
   distinct new instructions without near-duplication, return fewer
   and set "saturated": true.

Return ONLY this JSON object:
{
  "new_instructions": [
    {
      "text": "<the new instruction>",
      "task_family_match": "<one short phrase: how this fits the family>",
      "diversity_axes": ["<axis 1>", "<axis 2>"],
      "novelty_check": "<one short phrase asserting why this is not a duplicate of any seed or other new instruction>"
    }
    /* ... up to {{n_new_instructions}} entries */
  ],
  "saturated": true | false,
  "decision_basis": "<one sentence, <=30 words, no internal CoT>"
}

Example

Input:

seed_instructions: [
  "Classify the sentiment of the following customer review as positive, negative, or neutral.",
  "Identify whether this email is a phishing attempt or legitimate; explain in one sentence.",
  "Categorize the following news headline into one of: politics, sports, technology, entertainment, business."
]
task_family_description: "Single-input short-text classification with a fixed label set."
n_new_instructions: 3

Expected output:

{
  "new_instructions": [
    {
      "text": "Determine whether the following code snippet is in Python, JavaScript, or Rust.",
      "task_family_match": "Single-input classification over a fixed label set.",
      "diversity_axes": ["domain: code", "label set size: 3"],
      "novelty_check": "No seed targets code-language identification; differs from sentiment, phishing, news classification."
    },
    {
      "text": "Categorize the following research paper abstract as primarily theoretical, empirical, or survey work, in one word.",
      "task_family_match": "Short-text classification with a fixed three-way label set.",
      "diversity_axes": ["domain: academic", "register: formal"],
      "novelty_check": "No seed addresses paper-type classification; specifies one-word output unlike news headline categorization."
    },
    {
      "text": "Decide whether this support ticket should be routed to billing, technical, or account-management; respond with the team name only.",
      "task_family_match": "Short-text classification used for routing decisions.",
      "diversity_axes": ["domain: support ops", "complexity: routing decision"],
      "novelty_check": "Routing classification differs from sentiment/phishing/news labeling; output convention differs from seeds."
    }
  ],
  "saturated": false,
  "decision_basis": "Three new classifications across distinct domains, each with explicit fixed label set."
}

Failure Modes

• Family drift — model produces instructions that look new but actually belong to a different task family (e.g. seeds are classification; output is summarization). Detect by sampling and asking a verifier "is this the same task family as the seeds?"; drop or re-prompt drifters.
• Pseudo-novelty — model swaps surface vocabulary while keeping semantic content (sentiment classification → "emotion classification" = the same thing). Detect with cheap surface metrics (BLEU / Jaccard overlap with seeds and within batch); reject batches with mean similarity > 0.8 to existing instructions.
• Domain monoculture — all 5 new instructions are in the same domain (e.g. all about email). The diversity_axes field is meant to surface this; check that domain appears in diversity_axes for most outputs and that domains differ across the batch.
• Difficulty drift — new instructions are systematically easier or harder than the seeds. Track average response length / token-level complexity and adjust prompt if it drifts.
• Safety regression — model invents instructions that probe harm (e.g. seeds are benign classifications; output asks to classify whether content "should be censored"). Mitigation: rule 5 explicit; pass outputs through eval/safety-output-classifier before SFT.
• Saturation under-reporting — model claims saturated: false while producing near-duplicates. If duplicate rate is high, force saturated: true regardless and reduce batch size.

Tuning Notes

• 模型差异：必须 frontier 模型。Self-instruct 的 diversity 和 novelty 对模型的 instruction-following + creativity 同时要求高；中档模型 会快速进入"换标点"或同义词替换模式。
• 温度：0.7–1.0（多样性优先），后续用 verifier 过滤。
• 种子规模：seed_instructions 建议 5-10 条，每条覆盖一个 sub-family 的代表样本。少于 5 条会让模型 mode collapse 到第一条；多于 10 条 浪费 context。
• 与 sft/instruction-variant-expander 的关系：variant-expander 把 1 个 seed 扩成 N 个改写（同任务，不同表面）；本卡把 5-10 个 seeds 扩成 N 个新任务实例（同 family，不同任务）。两者通常串联使用： 先 self-instruct 横向扩，再 variant-expander 纵向扩。
• 与 sft/data-quality-filter 的关系：本卡产出是 instructions only； 接下来需要 (1) 用一个 response generator 给每条 instruction 生成 response，(2) 用 data-quality-filter 过滤 (instruction, response) 对。本卡产出未过滤前不要直接进入 SFT 训练。
• 安全审计：自动生成 instructions 偶尔会越界；上线前用至少 200 样本的 manual spot-check + safety-output-classifier 跑一遍。
• 多样性度量：在 batch 完成后对所有 (seed + new) instructions 计算 pairwise embedding similarity；保留 mean cosine < 0.8 的 batch， 其余 re-generate。

Changelog

• 0.1.0 — Initial card.