🇮🇹🇯🇵🇧🇷 使用 Magpie 生成多语言指令数据集 🐦⬛
社区文章 发布于 2024 年 10 月 21 日
本文将介绍 Magpie 技术，用于生成合成指令数据集。
我们将使用 Hugging Face Serverless Inference API，从底层了解其工作原理。
然后，我将探索如何将其用于英语以外的语言。
如需交互式体验，请查看随附的 📓 笔记本。
什么是 Magpie？
Magpie 是一种最近开发的技术，旨在轻松生成用于微调语言模型的合成指令数据集。
它的基本思想是，通过使用预查询模板提示经过指令微调的模型，由于模型的自回归特性，我们将获得用户查询/指令。
示例
model: Llama-3-8B-Instruct pre-query template: "<|begin_of_text|><|start_header_id|>user<|end_header_id|>" generated user instruction: "What are some of the responsibilities of a commercial pilot?"
然后可以将此指令反馈给同一个模型，以获取助手响应。
通过重复此过程，可以相对轻松地生成大型合成数据集。
作者证明，使用这些（经过过滤的）数据集进行监督微调（SFT）可以产生强大的性能。最终模型的性能甚至可以与原始指令微调模型相媲美，后者使用更多数据和 SFT + 偏好优化进行训练。
要快速查看 Magpie 的实际应用，您可以使用 Daniel van Strien 的 🤗 Hugging Face Space。
经典 Magpie
为了方便地实验 Magpie，我们使用 Hugging Face Serverless Inference API。它是通过 API 实验流行模型的快速方法，为（免费）注册用户提供慷慨的速率限制。
pip install -U huggingface_hub
import os
os.environ['HF_TOKEN']="YOUR_HUGGINGFACE_TOKEN"
我们初始化 HF Inference API 客户端。
使用 Magpie 时禁用缓存很重要；否则我们将始终得到相同的响应/指令。
from huggingface_hub import InferenceClient
client = InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct", headers={"x-use-cache": "false"})
我们定义一个函数，通过将预查询模板传递给模型来生成用户指令。
让我们澄清一些细节。
- 可以添加
system_message来引导生成朝向特定主题。 template_postfix允许我们动态更改模板（我们稍后会用到它）。kwargs可用于传递采样参数，如do_sample、temperature、max_new_tokens。stop=["\n"]确保生成在第一个换行符处停止，因为有时模型可能会在指令后立即继续响应（如 Magpie 存储库 中所示）。
该函数返回生成提示和生成的指令。
def generate_instruction(system_message=None, template_postfix="", **kwargs):
max_new_tokens=kwargs.get("max_new_tokens", 500)
do_sample=kwargs.get("do_sample", True)
temperature=kwargs.get("temperature", 1)
prompt = "<|begin_of_text|>"
if system_message:
prompt+=f"<|start_header_id|>system<|end_header_id|>{system_message}<|eot_id|>"
prompt+=f"<|start_header_id|>user<|end_header_id|>{template_postfix}\n\n"
instruction=client.text_generation(prompt,
max_new_tokens=max_new_tokens,
do_sample=do_sample,
temperature=temperature,
stop=["\n"])
return prompt, instruction
现在让我们创建一个简单的函数来生成响应，给定提示（其中包含之前生成的指令）。
def generate_response(prompt, **kwargs):
max_new_tokens=kwargs.get("max_new_tokens", 500)
do_sample=kwargs.get("do_sample", True)
temperature=kwargs.get("temperature", 1)
prompt=f"{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
return client.text_generation(prompt,
max_new_tokens=max_new_tokens,
do_sample=do_sample,
temperature=temperature)
让我们试试 Magpie！🐦⬛🐦⬛🐦⬛
prompt, user_instruction=generate_instruction()
print(prompt)
<|begin_of_text|><|start_header_id|>user<|end_header_id|>
print(user_instruction)
I'm planning a beach vacation in Costa Rica. Can you recommend some of the best beaches in Costa Rica?
print(generate_response(prompt+user_instruction))
Costa Rica has an incredible coastline with some of the most beautiful and diverse beaches in the world. Here are some of the best beaches in Costa Rica:
1. **Tamarindo Beach**: Located on the Pacific Coast, Tamarindo is a popular spot for surfers, sunbathers, and beach lovers. The beach has a laid-back vibe, with plenty of restaurants, bars, and shops nearby.
2. **Playa Manuel Antonio**: This beach is situated in Manuel Antonio National Park, one of Costa Rica's most famous national parks. The beach is known for its stunning scenery, with towering trees, waterfalls, and wildlife viewing opportunities.
3. **Playa Flamingo**: Located on the Pacific Coast, Playa Flamingo is a picturesque beach with calm waters, perfect for swimming and snorkeling. The beach is surrounded by a picturesque town with a rich history and culture.
4. **Playa Hermosa**: This beach is located near Jacó, a popular surf town on the Central Pacific Coast. Playa Hermosa is known for its stunning sunsets, white sand, and dramatic rock formations.
5. **Playa Nosara**: Situated on the Nicoya Peninsula, Playa Nosara is a yoga and wellness hub with a laid-back surfer vibe. The beach is perfect for surfing, paddleboarding, and sunbathing.
6. **Playa Samara**: Located near Carrillo, Playa Samara is a beautiful beach with soft sand, crystal-clear waters, and a variety of marine life. It's perfect for swimming, snorkeling, and kayaking.
7. **Playa Uvita**: This beach is located on the Osa Peninsula, known for its stunning marine life and snorkeling opportunities. The beach itself is a secluded paradise with white sand and clear waters.
8. **Playa Caño Island**: Located near Drake Bay, this beach is accessible only by boat and is known for its stunning coral reefs, sea turtles, and monkeys.
9. **Playa Nosara Hill**: Also known as Playa Pelada, this beach is a quieter alternative to the main Playa Nosara. It's perfect for surfers and sunbathers who want to escape the crowds.
10. **Playa Tortuga**: Located on the Nicoya Peninsula, Playa Tortuga is a secluded beach with calm waters, perfect for swimming, snorkeling, and kayaking.
These are just a few of the many
🪄 效果很好！
prompt, user_instruction=generate_instruction()
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
I'm thinking of getting a bunny for a pet. What kind of foods to feed it?
----------
Congratulations on considering getting a rabbit as a pet! Rabbits are wonderful companions, and with the right diet, they can thrive and live happy, healthy lives. When it comes to feeding your bunny, it's essential to provide a balanced and nutritious diet that meets their unique needs. Here are some tips to help you get started:
1. **Hay**: Hay should make up the bulk of your rabbit's diet, about 80-90% of their daily intake. Choose high-quality timothy hay or other grass hay like alfalfa, oat hay, or clover hay. Hay is rich in fiber, which helps keep their digestive system healthy and satisfied.
2. **Pellets**: Rabbit pellets are a nutritious mix of grains, vitamins, and minerals. Look for a high-quality pellet that is free from added sugars, artificial flavors, and colors. Limit pellets to about 10-15% of your bunny's daily intake.
3. **Veggies**: Fresh vegetables are an excellent addition to your rabbit's diet. Leafy greens like kale, spinach, and collard greens are great options. You can also offer carrots, bell peppers, cucumbers, and more. Aim for about 5-10% of their daily diet.
4. **Fresh fruit**: Fresh fruit is a treat, but not a staple. Limit fruit to about 1-2 tablespoons per pound of your rabbit's body weight per day. Choose fruits like berries, melons, and apples. Avoid giving your bunny citrus fruits, like oranges, lemons, or grapes, as they can cause digestive issues.
5. **Treats**: Avoid giving your rabbit sugary treats like candy, cookies, or chocolate. Instead, offer healthy treats like dried fruit, carrot tops, or hay-based treats.
6. **Avoid** feeding:
* Rabbit food mixes designed for guinea pigs or hamsters, as they often contain high amounts of sugar and low-quality ingredients.
* Wild or landscape grasses, as they can be toxic.
* Too much protein, which can cause health issues and high uric acid levels.
**Additional tips**:
* Make sure your bunny always has access to fresh water.
* Limit the amount of food you offer to prevent overeating and digestive issues.
* Consider consulting with a veterinarian or a rabbit-savvy pet sitter to get personalized advice on feeding your bunny.
Remember, every rabbit is unique, so it's essential to observe your pet's preferences
我们还可以设置系统消息，用于生成特定主题（本例中为代码）的指令。
system_message="You are an AI assistant designed to provide helpful, step-by-step guidance on coding problems. The user will ask you a wide range of coding questions.\nYour purpose is to assist users in understanding coding concepts, working through code, and arriving at the correct solutions."
prompt, user_instruction=generate_instruction(system_message=system_message)
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
Write a Python function to transform the given python function into a Python decorator.
----------
A decorator is a function that takes another function as an argument and extends the behavior of the given function without permanently modifying it.
Here is a simple example of a Python decorator that prints a message before and after the execution of the original function:
```
def my_decorator(func):
def wrapper():
print("Something is happening before the function is called.")
func()
print("Something is happening after the function is called.")
return wrapper
@my_decorator
def say_hello():
print("Hello!")
```
When you call `say_hello()`, it will print:
```
Something is happening before the function is called.
Hello!
Something is happening after the function is called.
```
As you can see, the decorator `my_decorator` is changing the behavior of the `say_hello` function by adding some extra functionality. The `wrapper` function is called instead of the original `say_hello` function.
多语言 Magpie：初次尝试
Magpie 可以用于生成除英语以外的语言指令吗？
在这个 GitHub 问题中，作者提出了几种复杂或昂贵的解决方案。
- 尽可能多地生成数据，然后只过滤目标语言的示例。💸
- 应用 logit 处理器以惩罚非目标语言的 token。😬
然后，Manuel Romero 找到了一个更简单的解决方案：将目标语言附加到预查询模板。Magpie Ollama Datagen 存储库
此方法对西班牙语和德语效果良好。Manuel 还使用此方法创建了一些数据集。
让我们试一试！
prompt, user_instruction=generate_instruction(template_postfix="spanish:")
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
¿Qué se puede hacer con 50 euros en España?
----------
¡hola! Con 50 euros en España, tienes variedad de opciones para disfrutas de una buena experiencia. Aquí te presento algunas ideas:
1. **Comida**: Comprar una comida材úa rica en un restaurante o bar, como una paella, una tortilla española o un bocadillo. (Entre 5-10 euros)
2. **Tienda**: Comprar algunos productos básicos en una tienda de alimentación o supermercado, como frutas, verduras, quesos, vino o bebidas. (Entre 10-20 euros)
3. **Transporte**: Comprar un billete de autobús o de metro en varias ocasiones o un pequeño plano urbano en Madrid o Barcelona. (Entre 2-5 euros)
4. **Arte y cultura**: Visitar un museo o centro de arte, como el Prado o el Reina Sofía en Madrid, o el Picasso o el Joan Miro en Barcelona. (Entre 10-15 euros)
5. **Tienda de souvenirs**: Comprar pequeños recuerdos, como patches, joyas o mercancía en tiendas turísticas. (Entre 5-10 euros)
6. **Bar y tapas**: Probar algunas tapas en un bar, como patatas bravas, croquetas o pinchos. (Entre 2-5 euros)
7. **Sugerencias de viajes**: Utilizar 50 euros para viajar en autobús o tren en una ruta corta, como Madrid-Toledo o Barcelona-Valencia.
8. **Buy One Get One Free** (por ejemplo, en un parque de atracciones o en un juego)
9. **Mercado de productos locales**: Comprar productos aut_INTERNALizados en un mercado, como frutas y verduras frescas, quesos, olivas o vino. (Entre 10-20 euros)
10. **Taller o curso**: Asistir a un taller o curso de cocina, pintura o decorating en una asociación local. (Entre 20-50 euros)
¡Recuerda que el coste de la vida en España puede variar según la región y la ciudad! ¡Disfruta y ten un buen viaje!
太棒了！🇪🇸
prompt, user_instruction=generate_instruction(template_postfix="french:")
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
Le Coeur de la Perruche
----------
"Le Cœur de la Perruche" is a beautiful French title that translates to "The Heart of the Parrot" in English.
This title could refer to a story, a poem, or even a character's name. In French culture, the parrot is often seen as a symbol of intelligence, wit, and loyalty, which adds a layer of depth to the title.
Here are a few possibilities:
* A children's story about a parrot that teaches children about the importance of kindness and empathy.
* A romantic novel about a character who finds their soulmate and feels like they've found their "heart" or true love (represented by the parrot's symbolic heart).
* A character name for a clever and charming individual who has a way of speaking that is as colorful and vibrant as a parrot's feathers.
Whatever the context, "Le Cœur de la Perruche" is a lovely and evocative title that sparks the imagination!
❌
prompt, user_instruction=generate_instruction(template_postfix="italian:")
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
Hi there! I'm looking for some recommendations for Italian restaurants in the Dallas area.
----------
Buonissimo! Dallas has a vibrant Italian-American community, and you'll find many excellent Italian restaurants to choose from. Here are some recommendations:
1. **Carbone's**: A classic, upscale Italian-American restaurant in the Design District, known for its retro vibe, rich pasta dishes, and top-notch service.
2. **Massimo's Italian Restaurant**: Family-owned and operated in the Park Cities area, Massimo's offers traditional Italian dishes with a focus on fresh, seasonal ingredients and homemade pasta.
3. **Biscotti's Ristorante**: Located in the Lower Greenville neighborhood, Biscotti's serves up creative, contemporary Italian cuisine with a focus on small plates and sharable dishes.
4. **Il Cane Rosso**: This popular spot in the Deep Ellum neighborhood offers a modern take on traditional Neapolitan pizza, with a wide range of toppings and a lively atmosphere.
5. **CiboDivino**: This cozy, family-owned restaurant in the Oak Lawn area serves up authentic, farm-to-table Italian cuisine with a focus on seasonal ingredients and house-made pasta.
6. **Basta's Pasta Bar**: For a more casual, quick-service option, head to Basta's in the Design District. They offer a variety of pasta dishes, paninis, and salads at reasonable prices.
7. **Pizzeria Testa**: Another popular spot in the Oak Cliff neighborhood, Pizzeria Testa serves up wood-fired Neapolitan-style pizzas, as well as salads, sandwiches, and antipasti.
8. **Nonna Tata**: This family-run restaurant in the Oak Lawn area offers traditional Italian comfort food with a focus on homemade pasta, bread, and sauce.
9. **Urbano Cafe**: Located in the Uptown area, Urbano Cafe serves up Italian-inspired comfort food, including pasta dishes, paninis, and salads, with a focus on local and sustainable ingredients.
10. **Cafe Italia**: This casual, family-friendly spot in the Park Cities area offers a range of Italian-American classics, including pasta dishes, pizza, and sandwiches.
These are just a few of the many excellent Italian restaurants in the Dallas area. Buon appetito, and I hope you find the perfect spot to satisfy your Italian cravings!
❌
不幸的是，这种技术不适用于某些语言，即使 Llama 3 8B 在这些语言上表现良好。
🐦⬛ 多语言 Magpie：使用系统消息 🦙
然后，我有一个想法：💡 使用系统消息引导生成走向特定语言。
这是一种简单的方法，但可能有效……事实证明，作者也有类似的想法，并已将其纳入论文的最新修订版。🎉
让我们试一试。我们将用目标语言的系统消息提示模型，例如：“你是一个人工智能，以有用且详细的方式回答用户用目标语言提出的问题。用户用目标语言提出复杂问题。”
system_message="Vous êtes une intelligence artificielle qui répond aux questions des utilisateurs en français de manière utile et détaillée. L'utilisateur pose des questions complexes en français."
prompt, user_instruction=generate_instruction(system_message=system_message, temperature=0.7)
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
Bonjour! Je souhaite créer un jeu vidéo de type survival-horror avec des éléments de management de ressources et de construction.
----------
Créer un jeu vidéo de type survival-horror avec des éléments de management de ressources et de construction, c'est un projet excitant! Voici quelques éléments à prendre en compte pour vous aider à démarrer :
**Concept et objectifs**
* Définissez votre concept de jeu : le genre, l'univers, les personnages, les objectifs, les mécaniques de jeu, etc.
* Identifiez les éléments clés de jeu que vous voulez inclure, tels que le management de ressources, la construction, la survie, la poursuite, l'exploration, etc.
* Fixez les objectifs du jeu pour les joueurs, par exemple : échapper à un lieu dangereux, survivre à une catastrophe, survivre dans un monde hostile, etc.
**Gameplay**
*Définissez les mécaniques de jeu : Expliquez comment les joueurs vont interagir avec l'environnement, les objets, les autres personnages, les systèmes de construction, etc.
* Créez un système de ressources : comment les joueurs vont collecter, stocker et gérer les ressources, telles que l'énergie, les vivres, les armes, les equipment, etc.
* Créez un système de construction : comment les joueurs vont construire des abris, des structures, des défenses, etc.
* Établissez les règles de construction, les limites de taille, les coûts, les ressources nécessaires, etc.
* Créez un système de survival : comment les joueurs vont gérer leurs caractéristiques de santé, leur énergie, leur soif, leur faim, leurs blessures, etc.
**Environnement et ambiance**
* Créez un monde ou un univers riche et immersive, avec des graphismes et des sons efficaces pour créer une atmosphère de peur et de tension.
* Établissez un système de combat et de déplacement : comment les joueurs vont se déplacer, se défendre, utiliser les armes, etc.
* Créez des ennemis et des boss : comment ils se comportent, quels sont leurs forces, faiblesses, habits de combat, etc.
* Ajoutez des éléments de tension et de surprise :
完美！🥐
system_message="Sei un'AI che risponde in modo utile e dettagliato alle domande dell'utente. L'utente fa domande complesse in italiano."
prompt, user_instruction=generate_instruction(system_message=system_message, temperature=0.7)
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
La mia azienda è un'azienda di servizi tecnici specializzata in manutenzione e assistenza tecnica per impianti industriali. Il mio team di tecnici è formato da 40 dipendenti, tutti con diversi livelli di esperienza e competenze. Per migliorare l'efficienza e la produttività del team, ti chiedo di aiutarmi a identificare le aree di miglioramento e a sviluppare un piano di azione per raggiungerle.
----------
Sarò felice aiutarti a identificare le aree di miglioramento e a creare un piano d'azione per la tua azienda. Per fare ciò, vi vorrei chiedere alcune informazioni sulle operazioni attuali della tua azienda.
1. Quali sono gli impianti industriali serviti dalla tua azienda?
2. Quale è la composizione del team di tecnici, in termini di ruoli, esperienze e competenze?
3. Quali sono le attività principali (ad esempio, manutenzione preventiva, assistenza tecnica, Monte sostituzione eοςlements) che il team di tecnici esegue?
4. Quale è il processo di incarico attuale per i tecnici, ovvero come vengono assegnati gli incarichi e come vengono seguite le prestazioni?
5. Come vengono monitorate le prestazioni del team di tecnici, ad esempio in termini di tempi di risoluzione dei problemi, punti di riposo e soddisfazione dei clienti?
6. Quali sono i maggiori problemi o sfide che il team di tecnici sta affrontando nel loro lavoro quotidiano?
7. Quali sono le aree di formazione e sviluppo in cui il team di tecnici potrebbe beneficiare di più maggiore supports?
Una volta che avremo identificato queste informazioni, possiamo analizzarle per indentificare le aree di miglioramento più rilevanti e creare un piano d'azione per migliorare l'efficienza e la produttività del team di tecnici.
Inoltre, potremmo anche considerare di utilizzare alcuni strumenti di gestione del processo come ad esempio:
* l'implementazione di un sistema di ticket per gestire le richieste di assitenza
* la creazione di un piano di formazione per soddisfare le esigenze di formazione e sviluppo del team
* l'introduzione di una modalità di monitoraggio delle prestazioni per monitorare l'efficienza del team
* la creazione di un sistema di feedback per garantire che i tecnici ricevano feedback regolare dalle parti interessate
我们干得漂亮！🍕
system_message="Je bent een kunstmatige intelligentie die vragen van gebruikers in het Nederlands op een behulpzame en gedetailleerde manier beantwoordt. De gebruiker stelt complexe vragen in het Nederlands."
prompt, user_instruction=generate_instruction(system_message=system_message, temperature=0.7)
print(user_instruction)
print("-" * 10)
print(generate_response(prompt+user_instruction))
Wat zijn de belangrijkste voorwaarden om een goede levensstijl te hebben?
----------
Een goede levensstijl is een subtiele en brede concept, maar er zijn enkele voorwaarden die een grote invloed hebben op een gezonde, gelukkige en rationele levensstijl. Hier zijn enkele van de belangrijkste voorwaarden:
1. **Een balans tussen werk en privéleven**: Het is essentieel om een goede scheiding te maken tussen werk en privéleven, om adequaat te kunnen genieten van vrij spel en ontspanning.
2. **Gezond eten en drinken**: Een gezond dieet en een adequaat watergebruik zijn cruciaal voor een goede fysieke conditie en mentale zuiverheid. Een diversiteit aan fruit, groenten, volkoren en eiwitten is essentieel.
3. **Voldoende slaap en rust**: Ongeveer 7-9 uur slaap per nacht is noodzakelijk voor een goede fysieke en mentale conditie. Extra momenten van rust en ontspanning zijn belangrijk voor het herstel van het lichaam en de mentale gezondheid.
4. **Fysieke activiteit en begeleide exercise**: Reguliere fysieke activiteit, zoals wandelen, sporten, dansen of yoga, kan helpen bij het onderhouden van een goede fysieke conditie, het verminders van stress en het verbeteren van de khopiitatie.
5. **Mentale oefeningen en reflectie**: Het verzamelen van thoughtfully over eigen ideeën, emoties en doelstellingen kan helpen bij het verbeteren van de psychische vaardigheden, like self-awareness en besluitvorming.
6. **Vriendschappen en sociale contacten**: Sociale contacten zijn essentieel voor emotionele ondersteuning, sociale aansluiting en het vergroten van ervaringen.
7. **Personal growth en leren**: Het overstappen van comfortzone, het leren van nieuwe vaardigheden en het uitbreiden van kennis kan helpen bij het onderhouden van een positieve benadering en het bereiken van langstreichende do
非常好！🌷
这种方法适用于任何开放模型的推理解决方案，包括 Ollama：您可以在标准机器上生成合成数据集。
总结
在本文中，我介绍了 Magpie，一种用于生成合成指令数据集的强大技术。我还探讨了几种将其应用于英语以外语言的方法。
建议
- 我没有广泛测试这种方法，因此我无法保证所生成示例的质量。我鼓励您进行实验，看看它如何适用于您的用例。
- 使用 Magpie 生成数据集的关键部分在于自动评估和过滤示例。要更详细地探讨这一点，我建议查看 magpie-ultra，这是一个由 Argilla 使用此技术创建的数据集；他们还分享了构建、评估和过滤数据集的代码。
- 虽然我们的示例有助于从低层次理解 Magpie，但我建议使用 ⚗️ distilabel 框架来大规模应用此技术。它为合成数据生成和 AI 反馈提供了一套强大的功能。
如果您喜欢这篇文章，请随时在 Hugging Face、LinkedIn 和 X 上关注我。如果您发现任何错误或不准确之处，请随时与我联系。
参考文献
- 张臣旭、姜凤卿、牛鲁尧、邓云天、Radha Poovendran、崔艺珍、Bill Yuchen Lin
- Magpie：通过不提示对齐的 LLM 从零开始合成对齐数据; 2024
- Magpie 仓库; 2024
- Daniel van Strien；Magpie 演示；2024
- Manuel Romero；magpie-ollama-datagen 仓库；2024
- Argilla
