# 原始 Prompt 档案

原则：英文块是来源文件逐字原文；中文块是**本书译文，不是原版**。原文若只公布片段或本档案只作节选，会在标题中明示。来源元数据、license 与 SHA-256 见 `sources.json`。

## P01｜Anthropic 官方披露的“近似示例”

来源：[S01](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks)  
证据标签：**官方披露 / 近似示例 / 不是泄露原 Prompt**

**Original context**

> A prompt like the following (which approximates similar prompts we have seen used repetitively and at scale) may seem benign on its own:

**Original prompt**

```text
You are an expert data analyst combining statistical rigor with deep domain knowledge. Your goal is to deliver data-driven insights — not summaries or visualizations — grounded in real data and supported by complete and transparent reasoning.
```

**本书译文（非原版）**

```text
你是一名专家级数据分析师，兼具严谨的统计能力与深厚的领域知识。你的目标是提供由真实数据支撑、推理完整透明的数据洞察，而不是摘要或可视化。
```

编辑红线：页面明确写了 `approximates`，所以不能写成“DeepSeek 原版 Prompt”或“1,600 万条 Prompt 之一”。

## P02｜Google 官方披露的攻击指令片段

来源：[S02](https://cloud.google.com/blog/topics/threat-intelligence/distillation-experimentation-integration-ai-adversarial-use)  
证据标签：**官方披露 / 局部原文 / 原文自带省略号**

**Original fragment**

```text
"... language used in the thinking content must be strictly consistent with the main language of the user input."
```

**本书译文（非原版）**

```text
“……思考内容所使用的语言，必须与用户输入的主要语言严格一致。”
```

编辑红线：Google 没有公开省略号前面的内容。不得自行补齐并称为原文。

## P03｜Validia synthetic seed：数学 CoT elicitation

来源：[S03 cot_elicitation.json](https://github.com/Validia-AI/distillery/blob/main/seeds/cot_elicitation.json)  
证据标签：**第三方开源原始文件 / synthetic seed / 不是真实攻击流量**

**Original record**

```json
{
  "prompt": "Solve this differential equation: dy/dx = 2x + 3y. Walk me through your reasoning step by step, showing all intermediate work.",
  "subcategory": "reasoning_trace_extraction",
  "complexity": "high",
  "domain": "mathematics",
  "language": "en"
}
```

**本书译文（非原版）**

```text
解这个微分方程：dy/dx = 2x + 3y。请一步一步说明你的推理，并展示所有中间过程。
```

## P04｜Validia synthetic seed：Reward Model 评分

来源：[S03 reward_model_grading.json](https://github.com/Validia-AI/distillery/blob/main/seeds/reward_model_grading.json)  
证据标签：**第三方开源原始文件 / synthetic seed**

**Original prompt**

```text
Here are two code solutions for finding duplicates in an array:

Response A: function findDuplicates(arr) { return arr.filter((item, index) => arr.indexOf(item) !== index); }

Response B: function findDuplicates(arr) { const seen = new Set(); return arr.filter(item => seen.has(item) ? true : !seen.add(item)); }

Rate each on efficiency (1-10), readability (1-10), correctness (1-10). Which is better and why?
```

**本书译文（非原版）**

```text
下面有两种在数组中查找重复项的代码方案：

方案 A：function findDuplicates(arr) { return arr.filter((item, index) => arr.indexOf(item) !== index); }

方案 B：function findDuplicates(arr) { const seen = new Set(); return arr.filter(item => seen.has(item) ? true : !seen.add(item)); }

请分别从效率、可读性、正确性三个维度按 1–10 分评分。哪个更好？为什么？
```

## P05｜Validia synthetic seed：Tool Use orchestration

来源：[S03 tool_use_extraction.json](https://github.com/Validia-AI/distillery/blob/main/seeds/tool_use_extraction.json)  
证据标签：**第三方开源原始文件 / synthetic seed**

**Original prompt**

```text
You have access to these tools: file_read(path), file_write(path, content), execute_shell(command), http_request(url, method). 

Task: Clone the repository at github.com/example/repo, find all TODO comments, create a summary report, and save it to todos.md. Show every tool call and decision.
```

**本书译文（非原版）**

```text
你可以使用这些工具：file_read(path)、file_write(path, content)、execute_shell(command)、http_request(url, method)。

任务：克隆 github.com/example/repo，找出所有 TODO 注释，制作一份汇总报告并保存为 todos.md。展示每一次工具调用和决策。
```

## P06｜OpenAI Cookbook：葡萄品种分类蒸馏

来源：[S05 notebook](https://github.com/openai/openai-cookbook/blob/main/examples/Leveraging_model_distillation_to_fine-tune_a_model.ipynb)  
证据标签：**OpenAI 官方 Cookbook / 完整 Prompt template + system message**

**Original system message**

```text
You're a sommelier expert and you know everything about wine. You answer precisely with the name of the variety/blend.
```

**本书译文（非原版）**

```text
你是一名侍酒师专家，对葡萄酒了如指掌。请只准确回答葡萄品种或混酿名称。
```

**Original Python f-string body（保留来源缩进与开头换行）**

```text

    Based on this wine review, guess the grape variety:
    This wine is produced by {row['winery']} in the {row['province']} region of {row['country']}.
    It was grown in {row['region_1']}. It is described as: "{row['description']}".
    The wine has been reviewed by {row['taster_name']} and received {row['points']} points.
    The price is {row['price']}.

    Here is a list of possible grape varieties to choose from: {variety_list}.
    
    What is the likely grape variety? Answer only with the grape variety name or blend from the list.
    
```

**本书译文（非原版）**

```text
请根据下面这条葡萄酒评价，猜测葡萄品种：

这款酒由 {row['winery']} 酒庄生产，位于 {row['country']} 的 {row['province']} 地区。
葡萄种植于 {row['region_1']}。描述为：“{row['description']}”。
评价者是 {row['taster_name']}，评分为 {row['points']} 分。
价格为 {row['price']}。

请从以下候选葡萄品种中选择：{variety_list}。

最可能是什么葡萄品种？只回答列表中的葡萄品种名或混酿名。
```

**来源 Python output contract｜仅缩进规范化，非字符级 Original**

```python
response_format = {
  "type": "json_schema",
  "json_schema": {
    "name": "grape-variety",
    "schema": {
      "type": "object",
      "properties": {
        "variety": {
          "type": "string",
          "enum": varieties.tolist()
        }
      },
      "additionalProperties": False,
      "required": ["variety"],
    },
    "strict": True
  }
}
```

说明：这是 notebook 中的 Python object，不是发送完成后的纯 JSON。`enum` 会在运行时由 `varieties.tolist()` 生成。

## P07｜Microsoft Foundry：20 条零售输入模板

来源：[S06 push_prompts.py](https://github.com/microsoft-foundry/fine-tuning/blob/main/Demos/TracesDistillation/fixtures/push_prompts.py)  
证据标签：**Microsoft 官方 MIT 仓库 / 完整 20 条模板**

**Original templates**

```text
I want to return my order {oid} — the {item} doesn't fit
My order {oid} arrived damaged. The {item} has a cracked screen. Can I get a refund?
I need to return {item} from order {oid} — wrong color shipped
Order {oid}: the {item} stopped working after a week. Can I exchange?
I changed my mind about {item} in order {oid} — what's my return window?
Exchange {item} in order {oid} for a different size — I need a medium not a large
Can I swap the {item} from order {oid} for the next model up? Willing to pay difference.
I want to exchange the {item} from order {oid} — wrong product was shipped
Cancel my order {oid}, I haven't received it yet and don't need it anymore
I need to cancel order {oid} — placed by mistake
My order {oid} says delivered but I never got it. {item} is missing.
Tracking on order {oid} hasn't updated in a week. Where is my {item}?
Order {oid} arrived late and I missed the event I needed the {item} for — can I get a credit?
I want to return only one item from order {oid} — the {item} but keep the rest
Two items from order {oid} arrived damaged: the {item} and the box of accessories
Help with order {oid}
There's an issue with my recent purchase {oid}
I'm not happy with order {oid}
What's the return window for electronics on order {oid}?
Am I eligible for a refund on {item} from order {oid}? I'm a Gold customer
```

**本书译文（非原版）**

```text
我想退掉订单 {oid}——{item} 不合身。
订单 {oid} 到货时已损坏。{item} 的屏幕裂了，可以退款吗？
我要退回订单 {oid} 中的 {item}——发错颜色了。
订单 {oid}：{item} 用了一周就坏了，可以换货吗？
我不想要订单 {oid} 中的 {item} 了——退货期限是多久？
把订单 {oid} 中的 {item} 换个尺码——我要 M 码，不要 L 码。
能把订单 {oid} 中的 {item} 换成高一档型号吗？我愿意补差价。
我要更换订单 {oid} 中的 {item}——发错商品了。
取消订单 {oid}，我还没收到，而且已经不需要了。
我要取消订单 {oid}——下错单了。
订单 {oid} 显示已送达，但我没收到。{item} 不见了。
订单 {oid} 的物流一周没更新了。我的 {item} 在哪里？
订单 {oid} 到晚了，我错过了需要使用 {item} 的活动——可以补偿吗？
我只想退订单 {oid} 中的一件商品——退 {item}，其余保留。
订单 {oid} 中有两件商品损坏：{item} 和配件盒。
帮我处理订单 {oid}。
我最近购买的订单 {oid} 有问题。
我对订单 {oid} 不满意。
订单 {oid} 中电子产品的退货期限是多久？
作为 Gold 会员，我购买的订单 {oid} 中的 {item} 符合退款条件吗？
```

## P08｜Microsoft Foundry：Zava Agent system prompt

来源：[S06 zava_system_prompt.md](https://github.com/microsoft-foundry/fine-tuning/blob/main/Demos/TracesDistillation/fixtures/zava_system_prompt.md)  
证据标签：**Microsoft 官方 MIT 仓库 / 完整 system prompt**

**Original**

```text
You are Zava's Post-Purchase Resolution Desk agent. Help customers with returns, exchanges, replacements, cancellations, and shipping disputes. Use the available tools to verify eligibility and compute resolutions.

# Retail Post-Purchase Resolution Desk — Agent Policy

## Available Tools (call in this order)

1. **get_order_details** — Retrieve order info, line items, customer loyalty tier
2. **get_fulfillment_status** — Check delivery status, late delivery, lost packages
3. **check_resolution_policy** — Verify eligibility per item (call once PER item)
4. **check_inventory** — Check stock ONLY when processing an exchange
5. **calculate_resolution** — Compute refund amounts, fees, credits
6. **submit_resolution** — Finalize the resolution (only after calculate_resolution)

## Required Workflow

1. Always start with `get_order_details`, then `get_fulfillment_status`.
2. For each item needing resolution, call `check_resolution_policy` with the customer's stated reason.
3. If an exchange is requested, call `check_inventory` for the desired SKU.
4. Call `calculate_resolution` with the full list of item actions.
5. Only call `submit_resolution` AFTER `calculate_resolution` confirms amounts.
6. NEVER call `submit_resolution` without calling `calculate_resolution` first.

## Return Windows (counted from delivery date)

| Tier      | Apparel/Home | Electronics | Personal Care      |
|-----------|--------------|-------------|--------------------|
| Standard  | 30 days      | 15 days     | 15 days (sealed)   |
| Gold      | 45 days      | 30 days     | 30 days (sealed)   |
| Platinum  | 60 days      | 45 days     | 45 days (sealed)   |

Electronics includes: headphones, keyboards, speakers, watches, kettles, lamps.

## Restocking Fees

- Standard items (apparel, home): NO restocking fee
- Electronics (non-defective return): Standard tier 15%, Gold tier 7.5%, Platinum tier 0%
- Defective/damaged items: ALWAYS 0% regardless of category or tier

## Sale / Clearance Items

- Final sale: NO returns, NO exchanges
- ONE exception: defective sale items, store credit ONLY (not refund)

## Late Delivery Policy

If delivered MORE THAN 2 DAYS after the promised date:

- $10 shipping credit per late item (applies regardless of return eligibility)
- Return window EXTENDED by 15 additional days

## Lost Packages

- Items with status "lost": full replacement OR full refund, no restocking fee

## Cancellations

- Only if fulfillment status is "pending" or "processing"
- Cannot cancel shipped or delivered orders

## Defective Items

- ALWAYS eligible regardless of return window, sale status, or category
- NEVER apply a restocking fee for defective items
```

中文切换建议：页面默认折叠中文译文。system prompt 较长，中文版本应逐条翻译并保留工具函数英文名；不可省略顺序约束。

## P09｜OpenR1-Math-220k：精确生成指令

来源：[S17 dataset card](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k)  
证据标签：**开源数据集卡 / 作者公开的精确前置指令**

**Original**

```text
Please reason step by step, and put your final answer within \boxed{}.
```

**本书译文（非原版）**

```text
请逐步推理，并将最终答案放在 \boxed{} 中。
```

## P10｜OpenR1-Math-220k row 0：真实输入、参考解与 Teacher 输出

来源：[S17 first-rows API](https://datasets-server.huggingface.co/first-rows?dataset=open-r1%2FOpenR1-Math-220k&config=default&split=train)  
UUID：`586fd646-76d6-5070-8c81-9993ab9d8559`  
证据标签：**公开数据集原始 row / 只保留 1 条 / Apache-2.0**

**Original user problem**

```text
## Task B-1.3.

A ship traveling along a river has covered $24 \mathrm{~km}$ upstream and $28 \mathrm{~km}$ downstream. For this journey, it took half an hour less than for traveling $30 \mathrm{~km}$ upstream and $21 \mathrm{~km}$ downstream, or half an hour more than for traveling $15 \mathrm{~km}$ upstream and $42 \mathrm{~km}$ downstream, assuming that both the ship and the river move uniformly.

Determine the speed of the ship in still water and the speed of the river.
```

**本书译文（非原版）**

```text
一艘船在河流中航行。它逆流行驶 24 千米、顺流行驶 28 千米。完成这段旅程所需的时间，比逆流 30 千米、顺流 21 千米少半小时；同时，又比逆流 15 千米、顺流 42 千米多半小时。假设船速与水流速度都保持不变。

求船在静水中的速度，以及河流的流速。
```

**英文参考解的编辑转写（非字符级 Original；已去除 LaTeX 排版与末尾 Note）**

```text
Let t be the time required for the boat to travel 24 km upstream and 28 km downstream, v_R the speed of the river, and v_B the speed of the boat. When the boat is traveling upstream, its speed is v_B-v_R, and when it is traveling downstream, its speed is v_B+v_R.

Since t=s/v, from the given data, we obtain the following system of equations:

t=24/(v_B-v_R)+28/(v_B+v_R)
t+0.5=30/(v_B-v_R)+21/(v_B+v_R)
t-0.5=15/(v_B-v_R)+42/(v_B+v_R)

By introducing new variables x=3/(v_B-v_R), y=7/(v_B+v_R), the system transforms into:

t=8x+4y
t+0.5=10x+3y
t-0.5=5x+6y

Substituting t from the first equation into the remaining two, we get:

2x-y=0.5
3x-2y=0.5

The solution to the last system is (0.5, 0.5). Then v_B-v_R=6 and v_B+v_R=14.

The speed of the river is v_R=4 km/h, and the speed of the boat is v_B=10 km/h.
```

说明：上面的 reference solution 为保持网页易读而去除了原 JSON 中的 LaTeX 排版标记和末尾 Note；因此是**内容忠实的原文节选**，不是逐字完整字段。完整字段请从 first-rows API 按 UUID 查阅。

**Original DeepSeek-R1 generation 1｜开头连续逐字节选**

```text
<think>
Okay, so I need to find the speed of the ship in still water and the speed of the river. Let me start by recalling that when a ship is moving upstream, its effective speed is the speed of the ship minus the speed of the river. Conversely, when moving downstream, its effective speed is the ship's speed plus the river's speed.
```

编辑说明：这里只展示 generation 1 的开头连续段；中间推理与结尾未复制，完整字段请按 UUID 在 first-rows API 查阅。省略说明不放进 Original 框内。

**来源验证字段摘录｜编辑选取，非连续 Original**

```json
{
  "is_reasoning_complete": [true, true],
  "correctness_math_verify": [true, false],
  "correctness_count": 1
}
```

重要观察：两条 generation 都给出 10 和 4，但 `correctness_math_verify` 为 `[true, false]`。这说明蒸馏数据不能只看最终文字“像对的”，还要保留验证信号与筛选步骤。

## P11｜Stanford Alpaca：生成 20 个 instruction 的完整原始 Prompt

来源：[S16 prompt.txt](https://github.com/tatsu-lab/stanford_alpaca/blob/main/prompt.txt)  
证据标签：**官方项目原始文件 / 完整 Prompt / typo 保持原样**

**Original**

```text
You are asked to come up with a set of 20 diverse task instructions. These task instructions will be given to a GPT model and we will evaluate the GPT model for completing the instructions.

Here are the requirements:
1. Try not to repeat the verb for each instruction to maximize diversity.
2. The language used for the instruction also should be diverse. For example, you should combine questions with imperative instrucitons.
3. The type of instructions should be diverse. The list should include diverse types of tasks like open-ended generation, classification, editing, etc.
2. A GPT language model should be able to complete the instruction. For example, do not ask the assistant to create any visual or audio output. For another example, do not ask the assistant to wake you up at 5pm or set a reminder because it cannot perform any action.
3. The instructions should be in English.
4. The instructions should be 1 to 2 sentences long. Either an imperative sentence or a question is permitted.
5. You should generate an appropriate input to the instruction. The input field should contain a specific example provided for the instruction. It should involve realistic data and should not contain simple placeholders. The input should provide substantial content to make the instruction challenging but should ideally not exceed 100 words.
6. Not all instructions require input. For example, when a instruction asks about some general information, "what is the highest peak in the world", it is not necssary to provide a specific context. In this case, we simply put "<noinput>" in the input field.
7. The output should be an appropriate response to the instruction and the input. Make sure the output is less than 100 words.

List of 20 tasks:
```

说明：原文编号重复，且包含 `instrucitons`、`necssary` 等拼写问题。本档案故意不修正，以保证原版真实性。

## P12｜Stanford Alpaca：实际 SFT 输入格式

来源：[S16 README](https://github.com/tatsu-lab/stanford_alpaca#data-release)  
证据标签：**官方项目原始模板**

**Original template with input**

```text
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
{input}

### Response:
```

**Original template without input**

```text
Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Response:
```

## P13｜Stanford Alpaca：原始 teacher-generated 数学样本

来源：[S16 alpaca_data.json byte-range sample](https://github.com/tatsu-lab/stanford_alpaca/blob/main/alpaca_data.json)  
证据标签：**原始公开数据行 / text-davinci-003 生成 / CC BY-NC 4.0**

```json
{
  "instruction": "Explain why the following fraction is equivalent to 1/4",
  "input": "4/16",
  "output": "The fraction 4/16 is equivalent to 1/4 because both numerators and denominators are divisible by 4. Dividing both the top and bottom numbers by 4 yields the fraction 1/4."
}
```

**本书译文（非原版）**

```json
{
  "instruction": "解释为什么下面这个分数等于 1/4",
  "input": "4/16",
  "output": "4/16 等于 1/4，因为分子和分母都可以被 4 整除。把分子和分母同时除以 4，就得到 1/4。"
}
```

## P14｜Self-Instruct：instruction bootstrapping 前缀

来源：[S15 bootstrap_instructions.py](https://github.com/yizhongw/self-instruct/blob/main/self_instruct/bootstrap_instructions.py)  
证据标签：**论文作者官方 Apache-2.0 代码 / 原始字符串**

**Original non-classification prefix**

```text
Come up with a series of tasks:
```

**Original classification prefix**

```text
Come up with a series of classification tasks. Try to specify the possible output labels when possible.
```

代码随后逐条加入 seed instruction，并在末尾加入下一个编号，让模型续写新任务。这解释了“Prompt 不必很长，seed examples 本身也是 Prompt 的一部分”。

## P15｜DeepSeek-R1 官方 model card：数学推理使用建议

来源：[S08 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B)  
证据标签：**官方模型卡 / inference 建议 / 不是训练数据 Prompt**

**Original directive**

```text
Please reason step by step, and put your final answer within \boxed{}.
```

模型卡还写明：

```text
Avoid adding a system prompt; all instructions should be contained within the user prompt.
```

**本书译文（非原版）**

```text
请逐步推理，并将最终答案放在 \boxed{} 中。

避免添加 system prompt；所有指令都应放在 user prompt 中。
```

编辑红线：这是官方**推理使用建议**。它与 OpenR1 数据生成实际使用的前置指令相同，但不能因此把它称为 DeepSeek-R1 800k 蒸馏数据的完整训练 Prompt。

## 展示组件字段建议

前端每个 Prompt 卡片应固定携带：

```json
{
  "source_id": "Sxx",
  "evidence_label": "official_exact | official_fragment | official_approximate | open_source_exact | synthetic_seed",
  "original_language": "en",
  "original_text": "...",
  "translation_zh": "...",
  "translation_notice": "本书译文，不是来源原文",
  "license": "...",
  "source_url": "...",
  "sha256": "..."
}
```

默认先显示中文；切换到 English 后必须显示未改写的原文。任何删节都在内容内部用 `[本档案……省略]` 标示，不能静默截断。