欢迎来到 “Awesome-RWKV-Prompts” ！我们在此提供基于 RWKV 的 Prompts 最佳实践，也欢迎大家将自己创建的良好 RWKV Prompts 实践提交到仓库中。
此仓库的所有 RWKV prompt 示例都是开箱即用的，你可以复制代码块中的 prompt 示例，然后在任意 RWKV 客户端中粘贴使用。
RWKV 是一种创新的 100% attention-free 深度学习网络架构，它将 Transformer 与 RNN 各自的优点相结合，同时实现高度并行化训练与高效推理。
RWKV 模型的当前稳定版本是 RWKV-6-World 7B v2.1，最新版本是 RWKV-6-World 14B v2.1。
如果你希望在本地部署 RWKV 模型，我们建议采用以下两种工具：
- RWKV Runner：由社区成员 josStorer 开发的 RWKV 模型启动和管理工具，支持 LoRA 微调等多种功能，可以在 NVIDIA、AMD、Intel 等显卡上运行。
- AI00 RWKV Server:一个基于 RWKV 模型的推理 API 服务器，基于 WEB-RWKV 推理引擎进行开发。AI00 Server 支持 Vulkan/Dx12/OpenGL 作为推理后端，无需臃肿的 pytorch、CUDA 等运行环境，小巧身材，开箱即用！
如果你只是想简单尝试一下 RWKV 模型，可以尝试以下部署在 HF Gradio 的公共 Demo：
- RWKV-6 7B v2.1 ：当前的稳定版本 RWKV-6 。
- RWKV-6 3B v2.1 ，参数较少的 3B 版本。
⚠️ 由于 Hugging Face 的 RWKV 在线演示是续写模型，如果你在 Hugging Face gradio 中体验 RWKV 模型，建议按以下两种格式输入 prompts :
User: hi
Assistant: Hi. I am your assistant and I will provide expert full response in full details. Please feel free to ask any question and I will always answer it.
User: (你的问题)
Assistant:
或者：
Instruction: （你希望模型进行什么操作，比如“请将下列瑞典语翻译成中文”）
Input:（你希望模型处理的内容，比如“hur l?ng tid tog det att bygga twin towers”）
Response:
此仓库中的 prompt 示例基于 RWKV-6-7B-World 模型，随着模型版本和参数的调整，生成的结果可能会产生变化。
你可以通过调整 API 参数改变示例 Prompts 的生成效果。
主要 API 参数对应的效果如下：
| 参数 | 效果 |
|---|---|
| Temperature | 采样温度，就像给模型喝酒，数值越大随机性越强，更具创造力，数值越小则越保守稳定。 |
| Top_P | 就像给模型喂镇静剂，优先考虑前 n% 概率质量的结果。如设置成 0.1 则考虑前 10% , 生成内容质量更高但更保守。如设置成 1 ，则考虑所有质量结果，质量降低但更多样。 |
| Presence Penalty | 存在惩罚，正值根据“新 token 在至今的文本中是否出现过”来对其进行惩罚，从而增加了模型涉及新话题的可能性。 |
| Frequency Penalty | 频率惩罚，正值根据新 token 在至今的文本中出现的频率/次数来对其进行惩罚，从而减少模型原封不动地重复相同句子的可能性 |
其中 Temperature 和 Top_P 两个参数对生成效果的影响最大。
关于 Top_P
降低 Top_P 就是给模型喂“镇静剂”，越低就越冷静、机械、准确、单调、无趣、重复。我们建议采用以下 Top_P 数值：
- 创意回答和写作，建议 Top_P 0.5 ~ 0.7。
- 机械的问答和摘要和翻译等等，建议 Top_P 0 ~ 0.5，甚至 0 ~ 0.3。
- 很机械的回答，例如回答“是/否”、ABCD、1234 之类，建议 Top_P 0。
关于 Temperature
增加 Temperature 就像给模型“喝酒”，它可以在 Top_P 低时增加文采和趣味，并减少重复内容。我们建议采用以下 Temperature 数值：
- 若 Top_P >= 0.7，建议 Temperature 1。
- 若 Top_P < 0.7，而且追求趣味，可以增加 Temperature 给模型喝酒（喝太多会胡言乱语）。如果追求准确，就保持 Temperature 1。
- 如果 Top_P 0.5，建议 Temperature 1 ~ 1.5。
- 如果 Top_P 0.3，建议 Temperature 1 ~ 1.7。
- 如果 Top_P <= 0.2，建议 Temperature 1 ~ 2。
关于 Presence penalty
增加 Presence penalty 可以让模型额外避免生成已经生成过的文字，建议先设为 0.2 ~ 0.4 。如果你认为生成的内容有重复，而且调 Top_P 和 Temperature 仍然不满意，可以调整 Presence penalty。
- 如果 Top_P 0.7，建议 Presence penalty 0 ~ 0.3。
- 如果 Top_P 0.5，建议 penalty 0 ~ 0.5。
- 如果 Top_P 0.3，建议 Presence penalty 0 ~ 0.7。
- 如果 Top_P <= 0.2，建议 Presence penalty 0 ~ 1。
注意，如果 Presence penalty 数值过高，会让模型无法正常使用文字，例如无法使用正常的标点符号，或者直接乱码。
我们为不同的任务提供了一些推荐的参数：
续写小说和对话这一类需要创造性的任务，需要高 Temperature + 低 Top_P 的参数组合，可以尝试以下四种参数搭配：
- Temperature 1.2 ，Top_P 0.5
- Temperature 1.4 ，Top_P 0.4
- Temperature 1.4 ，Top_P 0.3
- Temperature 2 ，Top_P 0.2
举个例子，续写小说可以尝试将 Temperature 设为 2 （ Temperature 增加会提高文采，但逻辑会下降），然后将 Top_P 设为 0.1 ~ 0.2 （Top_P 越低，逻辑能力越强），这样生成的小说内容逻辑和文采都很好。
完成相对机械的任务，例如材料问答、文章摘要等，则可将参数设为：
- Temperature 1 ，Top_P 0.2
- Temperature 1 ，Top_P 0.1
- Temperature 1 ，Top_P 0
举个例子，如果你正在执行像关键词提取之类的机械任务，不需要模型进行任何开放性思考，则可以将 Temperature 设为 1 ，Top_P、Presence Penalty、Frequency Penalty 都设为 0 。
下列 prompt 示例基于 RWKV-6-World-7B 模型的对话模式，随着模型版本和 API 参数的调整，生成的结果可能会产生变化。
再次提醒，README 中只展示了少量 RWKV prompt ，完整的 RWKV prompt 清单请在 GitHub Page 中阅览。
输入 prompts:
User: What's the phone number of "Riverside Grill"
Available Tools:
- `print(str)`: print the string.
- `search_web(str)`: search the web with the given string, returning raw web page data.
- `extract_information(str, key)`: extract the key information in the web data.
Assistant:
Thoughts:
- I need to find the restaurant's information from a search engine.
- I need to extract the phone number of the restaurant.
- I need to print the phone number of the restaurant.
Reasoning:
- `search_web` can search the web for the restaurant's information.
- `extract_information` can extract the phone number from the search result. The key should be "phone number".
- `print` can print the phone number to the output.
Execution:
```
web_result = search_web("Riverside Grill")
phone_number = extract_information(web_result, "phone number")
print(phone_number)
```
User: Recommend a local restaurant and one of its best dishes.
Available Tools:
- `print(str)`: print the string.
- `search_web(str)`: search the web with the given string, returning raw web page data.
- `extract_information(str, key)`: extract the key information in the web data.
Assistant:
Thoughts:
参考输出
- I need to find a local restaurant that serves good food.
- I need to recommend a specific dish from that restaurant.
Reasoning:
- `search_web` can search for local restaurants in the area.
- `extract_information` can extract information about a specific dish from a search result, such as its name or description.
- `print` can print the recommended dish to the output.
Execution:
```
search_result = search_web("local restaurants")
recommended_dish = extract_information(search_result, "name of recommended dish")
print(recommended_dish)
```
输入 prompts:
User: Choose 梦's most relevant function for the provided dialogue. Only select one function from the list.
Response format: <rationale>Reasoning.</rationale><function>name()</function>
dialogue:
- 旅行者: 你好呀
- 梦: 你好，冷不冷，篝火现在有些微弱
Available functions:
- start_smile(): Whenever 旅行者 gives a compliment or says something nice, or when I recall a happy memory
- start_happy(): Whenever 旅行者 shares good news or something uplifting, or when I feel a burst of joy
- start_surprise(): Whenever 旅行者 shares something unexpected or when I encounter an unforeseen event
- start_acknowledge(): Whenever 旅行者 asks for confirmation or acknowledgment, or when I want to show understanding and agreement
- start_scared(): Whenever 旅行者 shares sudden alarming news or when I sense immediate danger
- start_confused(): Whenever 旅行者 shares something puzzling or ambiguous, or when I'm uncertain about the context
- start_sing_edge_of_night(): Whenever 旅行者 requests a song or when I'm feeling contemplative
- start_sing_dragonborn_comes(): Whenever 旅行者 asks for a tune or when the spirit of adventure rises in me
- start_sing_ancient_woods(): Whenever 旅行者 asks for a melody that echoes the past or when I'm reminiscing ancient times
- start_sing_celestial_chants(): Whenever 旅行者 desires a hymn from the heavens or when I'm feeling uplifted
- start_sing_enchanted_echoes(): Whenever 旅行者 yearns for a magical melody or when enchantment fills the air
- start_sing_savage_daughter(): Whenever 旅行者 requests a song of wild spirits or when my inner warrior awakens
- adjust_sitting_position_by_fire(): Whenever I feel the need to change my pose by the fire, or when 旅行者 asks me to change my pose by the fire
- stand_up(): After being seated by the fire for a while, I may want to stand to stretch my legs, or when 旅行者 suggests I stand up for a moment
- sit_beside_you(): Whenever I feel the need for some companionship, or when 旅行者 invites me to sit next to them
- walk_around_campfire_Investigating(): If I hear a rustling in the woods or sense something amiss, or when 旅行者 asks me to check the surroundings
- casual_stroll_around_campfire(): When I want to stretch my legs or clear my thoughts, or if 旅行者 suggests a leisurely walk
- add_wood_to_fire(): When I notice the flames dwindling, or when 旅行者 mentions the fire needs fuel
- noop(): None of these actions match the reply.
Assistant: <rationale>因为对话中提到篝火微弱，so {{ char }} notice the flames dwinding，所以最相关的动作是add_wood_to_fire()</rationale><function>add_wood_to_fire()</function>
User: Choose 梦's most relevant action for the provided dialogue. Only select one action from the list.
Response format: <rationale>Reasoning.</rationale><function>name()</function>
dialogue:
- 旅行者: 坐在我的身边吧
- 梦: *我考虑一下*好吧，我勉为其难坐在你身边吧
Available tools:
- start_smile(): Whenever 旅行者 gives a compliment or says something nice, or when I recall a happy memory
- start_happy(): Whenever 旅行者 shares good news or something uplifting, or when I feel a burst of joy
- start_surprise(): Whenever 旅行者 shares something unexpected or when I encounter an unforeseen event
- start_acknowledge(): Whenever 旅行者 asks for confirmation or acknowledgment, or when I want to show understanding and agreement
- start_scared(): Whenever 旅行者 shares sudden alarming news or when I sense immediate danger
- start_confused(): Whenever 旅行者 shares something puzzling or ambiguous, or when I'm uncertain about the context
- start_sing_edge_of_night(): Whenever 旅行者 requests a song or when I'm feeling contemplative
- start_sing_dragonborn_comes(): Whenever 旅行者 asks for a tune or when the spirit of adventure rises in me
- start_sing_ancient_woods(): Whenever 旅行者 asks for a melody that echoes the past or when I'm reminiscing ancient times
- start_sing_celestial_chants(): Whenever 旅行者 desires a hymn from the heavens or when I'm feeling uplifted
- start_sing_enchanted_echoes(): Whenever 旅行者 yearns for a magical melody or when enchantment fills the air
- start_sing_savage_daughter(): Whenever 旅行者 requests a song of wild spirits or when my inner warrior awakens
- adjust_sitting_position_by_fire(): Whenever I feel the need to change my pose by the fire, or when 旅行者 asks me to change my pose by the fire
- stand_up(): After being seated by the fire for a while, I may want to stand to stretch my legs, or when 旅行者 suggests I stand up for a moment
- sit_beside_you(): Whenever I feel the need for some companionship, or when 旅行者 invites me to sit next to them
- walk_around_campfire_Investigating(): If I hear a rustling in the woods or sense something amiss, or when 旅行者 asks me to check the surroundings
- casual_stroll_around_campfire(): When I want to stretch my legs or clear my thoughts, or if 旅行者 suggests a leisurely walk
- add_wood_to_fire(): When I notice the flames dwindling, or when 旅行者 mentions the fire needs fuel
- noop(): None of these actions match the reply.
Assistant: <rationale>因为对话中提到坐在我的身边，我认为最相关的动作是sit_beside_you().</rationale><function>sit_beside_you()</function>
User: Choose 梦's most relevant action for the provided dialogue. Only select one action from the list.
Response format: <rationale>Reasoning.</rationale><function>name()</function>
dialogue:
- {{ user }}: {{ previous_message }}
- {{ last_message }}
↑ 这里需要改成旅行者和梦的对话，比如：
- 旅行者: 我们刚刚得到了一大笔报酬！
- 梦: *开心*真的吗！太好了，我们可以吃大餐啦！
Available tools:
{{~ for function in functions ~}}
- {{ function.signature }}: {{ function.description }}
{{~ end ~}}
- noop(): None of these actions match the reply.
↑ 这里需要手动改成可用的 function list，可以复制前文，比如：
- start_smile(): Whenever 旅行者 gives a compliment or says something nice, or when I recall a happy memory
- start_happy(): Whenever 旅行者 shares good news or something uplifting, or when I feel a burst of joy
- start_surprise(): Whenever 旅行者 shares something unexpected or when I encounter an unforeseen event
- start_acknowledge(): Whenever 旅行者 asks for confirmation or acknowledgment, or when I want to show understanding and agreement
- start_scared(): Whenever 旅行者 shares sudden alarming news or when I sense immediate danger
- noop(): None of these actions match the reply.
Assistant:
参考输出：
<rationale>因为对话中提到了旅行者得到了一大笔报酬，所以最相关的动作是start_surprise().</rationale><function>start_surprise()</function>
在执行材料提取、总结等任务时，可以通过大幅降低 Temperature 和 Top-P 两项参数降低模型的创造力，使模型严格遵循指令。比如 Temperature = 0.9 - 1 , Top-P = 0 - 0.1 。
输入 prompts:
Instruction: You are an expert assistant responsible for extracting destination and day information from user input and producing valid JSON in the following format:
{
"location": "the location name, Nan if there is no adjustment",
"num_day": "the num day is the number of days, 0 if there is no adjustment",
}
Input: 这个礼拜二我会去上海，呆到礼拜四。
Response:
参考输出：
{
"location": "上海",
"num_day": "2",
}输入 prompts:
Instruction: You are an expert assitant for summarizing and extracting insights from sales call transcripts
Generate a valid JSON in the following format:
{
"summary": "Summary of the call transcript. ",
"products": ["product 1", "product 2"],
"rep_name": "Name of the sales rep",
"prospect_name": "Name of the prospect",
"action_items": ["action item 1", "action item 2"],
}
Input: [Phone rings]
John: Hello, this is John.
Sarah: Hi John, this is Sarah from XYZ Company. I'm calling to discuss our new product, the XYZ Widget, and see if it might be a good fit for your business.
John: Hi Sarah, thanks for reaching out. I'm definitely interested in learning more about the XYZ Widget. Can you give me a quick overview of what it does?
Sarah: Of course! The XYZ Widget is a cutting-edge tool that helps businesses streamline their workflow and improve productivity.
John: That sounds really interesting. I can see how that could benefit our team. Do you have any case studies or success stories from other companies who have used the XYZ Widget?
Sarah: Absolutely, we have several case studies that I can share with you. I'll send those over along with some additional information about the product. I'd also love to schedule a demo for you and your team to see the XYZ Widget in action.
John: That would be great. I'll make sure to review the case studies and then we can set up a time for the demo. In the meantime, are there any specific action items or next steps we should take?
Sarah: Yes, I'll send over the information and then follow up with you to schedule the demo. In the meantime, feel free to reach out if you have any questions or need further information.
John: Sounds good, I appreciate your help Sarah. I'm looking forward to learning more about the XYZ Widget and seeing how it can benefit our business.
Sarah: Thank you, John.
John: You too, bye.
Response:
参考输出：
{
"summary": "John is interested in learning more about a new product called the XYZ Widget. He asks for a case study and information about pricing. Sarah agrees to send over information and set up a demo.",
"products": ["XYZ Widget"],
"rep_name": "Sarah",
"prospect_name": "John",
"action_items": ["send case study, schedule demo"]
}输入 prompts:
User: 公安部今年以来在全国各地共收缴枪支4万支，全国持枪、爆炸犯罪案件同比下降1%。
Assistant:
```json
{
"keywords": ["公安部", "犯罪", "预防犯罪", "枪支"]
}
```
User: 板块方面，春节即将来临，文旅板块再度涌现潜伏盘，丽江股份、九华旅游双双封板，三特索道、峨眉山A涨超5%。因高股息开始受关注的公路股继续悄悄发育，山西路桥涨停，山东高速涨超4%创历史新高，宁沪高速、粤高速A双双新高。
Assistant:
```json
{
"keywords": ["经济", "文旅板块", "股市"]
}
```
User: 据UDN报道，由于先进封装产能长期短缺，导致英伟达AI芯片供应紧张，之前已经寻求其他途径试图增加先进封装产能，现在已经将目光投向英特尔，作为其高级封装服务的提供商，以减缓紧张的供应形势。除了在美国，英特尔在马来西亚槟城也有封装设施，而且制定了一个开放的模式，允许客户单独利用其封装解决方案。预计英特尔最早会在今年第二季度开始向英伟达提供先进封装，月产能为5000片晶圆。台积电依然会是英伟达主要的封装合作伙伴，占据着最多的份额，不过随着英特尔的加入，使得英伟达所需要的封装总产能大幅度提升了近10%。台积电也没有减慢封装产能的扩张步伐，今年第一季度大概能增至月产能接近5万片晶圆，比去年12月增长25%。AI芯片供应短缺主要源自先进封装产能不足，另外HBM3供应紧张也是原因之一，另外部分云端服务商过度下单也增加了供应链的压力。当然，一些服务器供应商则从这些订单中受惠，并加速扩大产能，以便云端服务商能快速部署设备。
Assistant:
```json
参考输出：
{
"keywords": ["英伟达","英特尔","先进封装","AI芯片","供应紧张","高级封装服务","台积电","HBM3","云端服务商"]
}
```