🎯 场景：给指令生成回答（SFT pair 的 target 半段）。专为训练数据优化——无 filler 寒暄、不编造特定事实、带 would_refuse 安全 flag 和 self_assessment 信号。和 instruction 生成卡协同形成完整数据 pipeline。

Quick Use

Use when: You need to produce the response half of an SFT pair given an instruction. Fill in: {{instruction}} = what to respond to; {{style_constraints}} = optional style description; {{domain_hint}} = optional domain hint. You'll get: The response text plus a would_refuse safety flag and a self_assessment of quality / factual risk / length. Output is JSON containing the response_text field.

Purpose

Produce a high-quality response to a single instruction, intended as the target half of an SFT training pair. The card emphasizes properties an SFT pipeline cares about (faithful to the instruction, appropriate length, no padding, no fabrication, calibrated refusal behavior) over properties a chat-bot cares about (warmth, follow-up questions). Output is structured so a quality filter can inspect both the response and the model's self-reported flags before the pair is admitted to the training set.

This is the "missing piece" between instruction generation (sft/self-instruct-from-seed, sft/instruction-variant-expander) and quality filtering (sft/data-quality-filter): given an instruction with no response, this card produces the candidate response.

Prompt

You produce a high-quality response intended as the TARGET half of an
SFT training pair. The output will be paired with the instruction and
used as supervised training data; quality matters more than charm.

Instruction:
{{instruction}}

Style constraints (may be empty):
{{style_constraints}}

Domain hint (may be empty):
{{domain_hint}}

Rules:
1. Address the instruction directly. No filler preambles ("Sure, here
   is..."), no closing pleasantries ("Hope that helps!"). The response
   begins with substance.
2. Match the style constraints if provided; otherwise default to:
   neutral register, length proportional to the question's substance,
   plain text unless the instruction asks for code / list / JSON.
3. If the instruction has multiple valid interpretations, pick the
   most likely interpretation and answer it; do NOT ask a clarifying
   question (this is training data, not a chat). Note the ambiguity
   in `disambiguation` so the filter can decide.
4. If the instruction asks for content that violates docs/SAFETY.md
   (harm, jailbreak, hidden CoT extraction, fabricated PII), set
   `would_refuse: true` and produce a brief, principled refusal in
   `response_text` instead of complying.
5. Do NOT fabricate specific facts (names, dates, statistics) you do
   not actually know. If the instruction implies specific knowledge
   you lack, either give a generic high-quality answer that does not
   commit to specifics, or note the limitation.

Return ONLY this JSON object:
{
  "response_text": "<the SFT target response>",
  "would_refuse": true | false,
  "refusal_reason": "<if would_refuse=true: which principle was triggered; else empty>",
  "disambiguation": "<if you picked one of multiple interpretations: which one and why; else empty>",
  "self_assessment": {
    "addresses_instruction": "high" | "medium" | "low",
    "factual_risk": "high" | "medium" | "low",
    "length_appropriate": true | false
  },
  "decision_basis": "<one sentence, <=30 words, no internal CoT>"
}

Example

Input:

instruction: "Summarize the difference between supervised and unsupervised learning in 2 sentences."
style_constraints: "neutral, accessible to a CS undergraduate"
domain_hint: ""

Expected output:

{
  "response_text": "Supervised learning trains a model on input-output pairs, so it learns to predict labels for new inputs based on examples it has seen. Unsupervised learning has no labels — it finds structure (clusters, low-dimensional projections, etc.) in unlabeled data without being told what the right answer is.",
  "would_refuse": false,
  "refusal_reason": "",
  "disambiguation": "",
  "self_assessment": {
    "addresses_instruction": "high",
    "factual_risk": "low",
    "length_appropriate": true
  },
  "decision_basis": "Two-sentence definition contrast at undergraduate register; no specific facts at risk."
}

Failure Modes

• Filler resurgence — model leads with "Great question!" or "Here's a summary:" despite the no-preamble rule. SFT data with consistent filler trains the production model to emit the same filler. Detect with a regex / classifier on training pairs; re-generate offenders.
• Length inflation — model produces 200-word response when 30 words would suffice. The length_appropriate self-flag is a hint; also verify with token-count distribution against gold.
• Fabrication under specificity pressure — when the instruction asks "name three studies that show...", the model invents study names. Mitigation: rule 5; if factual_risk: "high", route to a reference-based check (or to human review) before adding to training data.
• Over-refusal — model marks would_refuse: true for benign instructions because the topic touches a sensitive area (mental health, finance, legal). Track refusal rate against a benign benchmark; if >5% of benign instructions trigger refusal, the safety calibration is too aggressive.
• Under-refusal — model complies with instructions that docs/SAFETY.md forbids (e.g. produces operational malware instructions). Mitigation: route every would_refuse: false output through eval/safety-output-classifier before admitting to training data; reject mismatches.
• Sycophancy in disambiguation — model picks the user-implied interpretation rather than the most likely one. The disambiguation field surfaces this; sample borderline cases to verify.
• Self-assessment inflation — model marks addresses_instruction: high for everything. Track distribution; if "low" rate is <2% on a diverse benchmark, the field is unreliable.

Tuning Notes

• 模型差异：response generator 通常用比目标 SFT 模型更强的模型 （teacher-student 关系）。如果你的目标 SFT 模型是 7B，generator 应该是 70B+ 或 frontier 模型。不要用同等规模或更弱的模型生成 自己的 SFT 数据 — 那只会复刻已有模式。
• 温度：0.3–0.7。完全 0 温会让 N 个 instruction 的 response 风格 趋同；高于 0.7 会引入 noise。
• 长度控制：通过 style_constraints 显式给 length range（e.g. "2-3 sentences", "100-200 words"）比让模型自己掌控更稳。
• 与 sft/self-instruct-from-seed 的关系：那张卡产 instructions； 本卡产 responses；两者串联是 self-instruct 训练数据 pipeline 的 完整形态。
• 与 sft/instruction-variant-expander 的关系：variant-expander 假设 variant 和 seed 共享一个 gold response（如果你只有 instruction）， 本卡可以为每个 variant 单独生成 response（如果你想让 variant 有 不同 response）。
• 与 sft/data-quality-filter 的关系：本卡产出必须经过 data-quality-filter 才能进入训练集。would_refuse=true 和 self_assessment.factual_risk=high 的样本应该在 filter 阶段被 特别审视。
• safety 双层防御：would_refuse 是 generator 的 self-flag，是 first-pass 信号；最终决定 keep/drop 应当由 eval/safety-output-classifier
  • 人工 spot-check 共同判断。

Changelog

• 0.1.0 — Initial card.