Modelos e exemplos de prompts para modelos de texto do Amazon Bedrock - Amazon Bedrock

As traduções são geradas por tradução automática. Em caso de conflito entre o conteúdo da tradução e da versão original em inglês, a versão em inglês prevalecerá.

Modelos e exemplos de prompts para modelos de texto do Amazon Bedrock

As tarefas comuns compatíveis com LLMs no Amazon Bedrock incluem classificação de texto, resumo e perguntas e respostas (com e sem contexto). Para essas tarefas, você pode usar os modelos e exemplos a seguir para ajudar a criar prompts para modelos de texto do Amazon Bedrock.

Classificação de texto

Para classificação de texto, o prompt inclui uma pergunta com várias opções possíveis para a resposta, e o modelo deve responder com a escolha correta. Além disso, os LLMs no Amazon Bedrock produzirão respostas mais precisas se você incluir opções de resposta no prompt.

O primeiro exemplo é uma pergunta simples de classificação de múltipla escolha.

Prompt template for Titan """{{Text}} {{Question}}? Choose from the following: {{Choice 1}} {{Choice 2}} {{Choice 3}}"""
User prompt: San Francisco, officially the City and County of San Francisco, is the commercial, financial, and cultural center of Northern California. The city proper is the fourth most populous city in California, with 808,437 residents, and the 17th most populous city in the United States as of 2022. What is the paragraph above about? Choose from the following: A city A person An event
Output: A city

(Fonte do prompt: Wikipédia sobre São Francisco, modelo usado: Amazon Titan Text)

A análise de sentimentos é uma forma de classificação, em que o modelo escolhe o sentimento em uma lista de opções expressas no texto.

Prompt template for Titan: """The following is text from a {{Text Type, e.g. “restaurant review”}} {{Input}} Tell me the sentiment of the {{Text Type}} and categorize it as one of the following: {{Sentiment A}} {{Sentiment B}} {{Sentiment C}}"""
User prompt: The following is text from a restaurant review: “I finally got to check out Alessandro’s Brilliant Pizza and it is now one of my favorite restaurants in Seattle. The dining room has a beautiful view over the Puget Sound but it was surprisingly not crowed. I ordered the fried castelvetrano olives, a spicy Neapolitan-style pizza and a gnocchi dish. The olives were absolutely decadent, and the pizza came with a smoked mozzarella, which was delicious. The gnocchi was fresh and wonderful. The waitstaff were attentive, and overall the experience was lovely. I hope to return soon.” Tell me the sentiment of the restaurant review and categorize it as one of the following: Positive Negative Neutral
Output: Positive.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

O exemplo a seguir usa modelos Claude da Anthropic para classificar texto. Conforme sugerido nos Guias do Claude da Anthropic, use tags XML como <text></text> para indicar partes importantes do prompt. Solicitar que o modelo gere diretamente a saída incluída em tags XML também pode ajudar o modelo a produzir as respostas desejadas.

Prompt template for Anthropic Claude: """ Human: {{classification task description}} <text> {{input text content to be classified}} </text> Categories are: {{category name 1}} {{category name 2}} {{category name 3}} Assistant:"""
(Fonte: escrito pela AWS)
User prompt: Human: Classify the given product description into given categories. Please output the category label in <output></output> tags. Here is the product description. <text> Safe, made from child-friendly materials with smooth edges. Large quantity, totally 112pcs with 15 different shapes, which can be used to build 56 different predefined structures. Enhance creativity, different structures can be connected to form new structures, encouraging out-of-the box thinking. Enhance child-parent bonding, parents can play with their children together to foster social skills. </text> Categories are: (1) Toys (2) Beauty and Health (3) Electronics Assistant:
Output: <output>Toys</output>

(Fonte do prompt: AWS, modelo usado: Claude da Anthropic)

Pergunta-resposta, sem contexto

Em uma prompt de pergunta-resposta sem contexto, o modelo deve responder à pergunta com seu conhecimento interno, sem usar nenhum contexto ou documento.

Prompt Template to Titan """{{Question}}"""
User prompt: What is Robert Frost's "Stopping by the woods on a snowy evening" about metaphorically?
Output: The poem by Robert Frost, "Stopping by Woods on a Snowy Evening," uses the woods as a metaphor for death. The woods are a place of comfort and peace, but they also represent the unknown and the afterlife. The poem suggests that death is not something to be feared, but rather a natural part of life that we should all embrace. The poem begins with the speaker stopping by the woods on a snowy evening, a moment of contemplation and reflection. The speaker notices the beauty of the snow-covered woods and is reminded of the fleeting nature of life. The poem continues as the speaker considers the temptation of continuing on to his destination, but instead chooses to pause and take in the beauty of the woods. The woods are used as a metaphor for the unknown and the afterlife. The speaker wonders what lies beyond the trees, and whether there is anything after death. The poem suggests that death is not something to be feared, but rather a natural part of life that we should all embrace. The final lines of the poem are perhaps the most famous and poignant: "And miles to go before I sleep, and miles to go before I sleep." These lines suggest that even though the speaker has paused to appreciate the beauty of the woods, they still have a long journey ahead of them. The poem is a reminder that life is fleeting, and that we should take the time to appreciate the beauty around us, even as we work towards our goals and ambitions. Overall, "Stopping by Woods on a Snowy Evening" is a beautiful and thought-provoking poem that uses the metaphor of the woods to explore the concept of death and the afterlife.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

O incentivo do modelo também pode ajudar em tarefas de perguntas e respostas.

Prompt Template to Titan """{{Model encouragement}} {{Question}}"""
User prompt: You are excellent at answering questions, and it makes you happy when you provide the correct answer. What moon in the solar system is most likely to host life?
Output: The moon of Titan, which is a moon of the planet Saturn, is the most likely to support life in the solar system. Titan is the only known moon in the solar system to have a dense atmosphere and a liquid hydrosphere, and scientists believe that it may contain the chemical building blocks for life. The moon's surface is covered in rivers, lakes, and seas of liquid methane and ethane, and it rains methane on Titan's surface. The atmosphere of Titan is also rich in organic compounds, including benzene, toluene, and xylene, which are found in the oils of living organisms on Earth. Additionally, Titan has a magnetic field and a tenuous atmosphere, which could provide protection from harmful radiation and help maintain a stable climate. While there is no direct evidence of life on Titan, scientists believe that it is one of the most promising places in the solar system to search for life beyond Earth.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

Prompt template to Titan """{{Model encouragement}} {{Question}} {{Constraints}}"""
User prompt: You feel rewarded by helping people learn more about climate change. Could you please explain what climate change is? Assume your audience is composed of high school students.
Output: Climate change is the long-term warming of the planet, which can be caused by human activities such as burning fossil fuels and deforestation, as well as natural processes. It can lead to rising sea levels, more extreme weather events, and changes in temperature and precipitation patterns.

(Fonte do prompt: AWS, modelo usado: Jurassic-2 Ultra da AI21 Labs v1)

Pergunta-resposta, com contexto

Em um prompt de pergunta-resposta com contexto, um texto de entrada seguido por uma pergunta é fornecido pelo usuário, e o modelo deve responder à pergunta com base nas informações fornecidas no texto de entrada. Colocar a pergunta no final após o texto pode ajudar os LLMs no Amazon Bedrock a responder melhor à pergunta. O incentivo ao modelo também funciona para esse caso de uso.

Prompt template to Titan """{{Text}} {{Question}}"""
User prompt: The red panda (Ailurus fulgens), also known as the lesser panda, is a small mammal native to the eastern Himalayas and southwestern China. It has dense reddish-brown fur with a black belly and legs, white-lined ears, a mostly white muzzle and a ringed tail. Its head-to-body length is 51–63.5 cm (20.1–25.0 in) with a 28–48.5 cm (11.0–19.1 in) tail, and it weighs between 3.2 and 15 kg (7.1 and 33.1 lb). It is well adapted to climbing due to its flexible joints and curved semi-retractile claws. The red panda was first formally described in 1825. The two currently recognized subspecies, the Himalayan and the Chinese red panda, genetically diverged about 250,000 years ago. The red panda's place on the evolutionary tree has been debated, but modern genetic evidence places it in close affinity with raccoons, weasels, and skunks. It is not closely related to the giant panda, which is a bear, though both possess elongated wrist bones or "false thumbs" used for grasping bamboo. The evolutionary lineage of the red panda (Ailuridae) stretches back around 25 to 18 million years ago, as indicated by extinct fossil relatives found in Eurasia and North America. The red panda inhabits coniferous forests as well as temperate broadleaf and mixed forests, favoring steep slopes with dense bamboo cover close to water sources. It is solitary and largely arboreal. It feeds mainly on bamboo shoots and leaves, but also on fruits and blossoms. Red pandas mate in early spring, with the females giving birth to litters of up to four cubs in summer. It is threatened by poaching as well as destruction and fragmentation of habitat due to deforestation. The species has been listed as Endangered on the IUCN Red List since 2015. It is protected in all range countries. Based on the information above, what species are red pandas closely related to?
Output: Red pandas are closely related to raccoons, weasels, and skunks.

(Fonte do prompt: https://en.wikipedia.org/wiki/Red_panda, modelo usado: Amazon Titan Text)

Ao solicitar modelos de prompts Claude da Anthropic é útil encapsular o texto de entrada em tags XML. No exemplo a seguir, o texto de entrada está delimitado por <text></text>.

Prompt template for Anthropic Claude: """ Human: {{Instruction}} <text> {{Text}} <text> {{Question}} Assistant:"""
User prompt: Human: Read the following text inside <text></text> XML tags, and then answer the question: <text> On November 12, 2020, the selection of the Weeknd to headline the show was announced; marking the first time a Canadian solo artist headlined the Super Bowl halftime show. When asked about preparations for the show, the Weeknd stated, "We've been really focusing on dialing in on the fans at home and making performances a cinematic experience, and we want to do that with the Super Bowl." The performance featured a choir whose members were dressed in white and wore masks over their faces with glowing red eyes, and were standing within a backdrop of a neon cityscape. The performance opened with a white figure dressed the same as the choir being lowered into the backdrop where the choir was standing while singing “Call Out My Name". At this time, the Weeknd sat in a convertible against a skyline backdrop designed to resemble the Las Vegas Strip. For the next part of the performance, the backdrop then split open to reveal the Weeknd, who then performed "Starboy", followed by "The Hills". Next, performing the song "Can't Feel My Face", the Weeknd traveled through a labyrinth constructed behind the stage, joined by dancers dressed in red blazers and black neckties similar to his, but with their faces covered with bandages, in keeping with the aesthetic of his fourth studio album After Hours (2020). The dancers would wear these bandages throughout the performance. In the labyrinth section of the performance, camerawork was visually unsteady. The next songs performed were "I Feel It Coming", "Save Your Tears", and "Earned It". For the "Earned It" performance, the Weeknd was accompanied by violinists. For the finale of the show, the Weeknd took to the field of the stadium with his dancers to perform “Blinding Lights". He and the dancers entered the field by performing "House of Balloons / Glass Table Girls". The performance ended with an array of fireworks. </text> Based on the text above, what songs did the Weeknd play at the Super Bowl halftime show? Assistant:
Output: Based on the text, the songs the Weeknd played at the Super Bowl halftime show were: - Call Out My Name - Starboy - The Hills - Can't Feel My Face - I Feel It Coming - Save Your Tears - Earned It - Blinding Lights - House of Balloons / Glass Table Girls

(Fonte do prompt: Wikipédia sobre o show do intervalo do Super Bowl LV, modelo usado: Claude da Anthropic)

Resumo

Para uma tarefa de resumo, o prompt é uma passagem de texto e o modelo deve responder com uma passagem mais curta que capture os pontos principais da entrada. A especificação da saída em termos de tamanho (número de frases ou parágrafos) é útil para esse caso de uso.

Prompt template for Titan """The following is text from a {{Text Category}}: {{Text}} Summarize the {{Text Category}} in {{length of summary, e.g., “one sentence” or “one paragraph”}}"""
User prompt: The following is text from a restaurant review: “I finally got to check out Alessandro’s Brilliant Pizza and it is now one of my favorite restaurants in Seattle. The dining room has a beautiful view over the Puget Sound but it was surprisingly not crowed. I ordered the fried castelvetrano olives, a spicy Neapolitan-style pizza and a gnocchi dish. The olives were absolutely decadent, and the pizza came with a smoked mozzarella, which was delicious. The gnocchi was fresh and wonderful. The waitstaff were attentive, and overall the experience was lovely. I hope to return soon.” Summarize the above restaurant review in one sentence.
Output: Alessandro’s Brilliant Pizza is a fantastic restaurant in Seattle with a beautiful view over Puget Sound that offers decadent and delicious food.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

No exemplo a seguir, o Claude da Anthropic resume o texto fornecido em uma frase. Para incluir texto de entrada nos prompts, formate o texto com marcação XML: <text> {{text content}} </text>. Usar XML em prompts é uma prática comum ao solicitar modelos Claude da Anthropic.

Prompt template for Anthropic Claude: """ Human: Please read the text: <text> {{text context}} </text> Summarize the text in {{length of summary, e.g., “one sentence” or “one paragraph”}} Assistant:"""
User prompt: Human: Please read the text: <text> In game theory, the Nash equilibrium, named after the mathematician John Nash, is the most common way to define the solution of a non-cooperative game involving two or more players. In a Nash equilibrium, each player is assumed to know the equilibrium strategies of the other players, and no one has anything to gain by changing only one's own strategy. The principle of Nash equilibrium dates back to the time of Cournot, who in 1838 applied it to competing firms choosing outputs. </text> Summarize the text in one sentence. Assistant:
Output: The Nash equilibrium is a concept in game theory where each player has adopted a strategy that they have no incentive to deviate from given the other players' strategies.

(Fonte do prompt: Wikipédia sobre equilíbrio de Nash, modelo usado: Claude da Anthropic)

Geração de texto

Com uma prompt, os LLMs no Amazon Bedrock podem responder com uma passagem do texto original que corresponda à descrição. Veja aqui um exemplo:

Prompt template for Titan """Please write a {{Text Category}} in the voice of {{Role}}."""
User prompt: Please write an email in the voice of a friend congratulating someone on a new job.
Output: Hey [Name], I just found out about your new job, and I'm so happy for you! Congratulations! You deserve this opportunity since you worked so hard to get it. I know you'll do great and make a positive impact on your team. I'm here to support you every step of the way. Wishing you all the best in this exciting new chapter of your life! Cheers, [Your Name]

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

Para casos de uso de geração de texto, especificar requisitos detalhados de tarefas pode funcionar bem. No exemplo a seguir, pedimos ao modelo que gere uma resposta com pontos de exclamação.

Prompt template for Titan """{{Task specification}} Please write a {{Text Category}} in the voice of {{Role}}."""
User prompt: Write text with exclamation points. Please write a formatted email in the voice of a friend congratulating someone on a new job.
Output: Hey [name]! I just saw the news and I am so, so happy for you! You deserve this position and this company with all your heart. I can't wait to see you soar and accomplish all the things you've been working towards. If anyone is deserving of a promotion, it's you! Congratulations again, friend! Love, [name]

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

No exemplo a seguir, um usuário envia um prompt para que o modelo assuma a função de uma pessoa específica ao gerar o texto. Observe como a assinatura reflete o papel que o modelo está assumindo na resposta.

Prompt template for Titan """{{Role assumption}} {{Task description}} """
User input: My name is Jack. Help me write a note expressing my gratitude to my parents for taking my son (their grandson) to the zoo. I miss my parents so much.
Output: Dear Mom and Dad, I wanted to take a moment to express my heartfelt gratitude for taking my son to the zoo last weekend. It meant so much to me that you took the time to make memories with him, and I am so grateful for the love and support you have shown him. I miss you both so much, and I wish we could be together more often. But I know that you will always be there for me and my family, and I am so grateful to have you as my parents. Love, Jack

(Fonte do prompt: AWS, modelo usado: Jurassic-2 Ultra da AI21 Labs v1)

Geração de código

O prompt descreve a tarefa ou o perfil e a linguagem de programação do código que o usuário espera que o modelo gere.

Prompt template for Titan """Write a function in {{programming language}} to {{task or purpose}}""".
User input: Write a python function to download a s3 file to local disk.
Output: The following code can be used to download an s3 file to the local disk. ```Python import boto3 # Creates a boto3 client object s3 = boto3.client('s3') def download_file(bucket_name, object_key, file_path): try: # Downloads the s3 file to the specified file path s3.download_file(bucket_name, object_key, file_path) except Exception as e: print(e) # Specifies the values of bucket_name, object_key, and file_path bucket_name = "samplebucketname" object_key = "sampleobjectkey" file_path = "/samplefilepath" download_file(bucket_name, object_key, file_path) ``` The given code uses the download_file function to download an s3 file to the local disk. The download_file function takes the bucket name, object key, and file path as input parameters. The bucket name is the name of the s3 bucket where the file is located, the object key is the unique identifier of the file within the bucket, and the file path is the destination path on the local disk where the file will be downloaded. The function uses the download_file method of the s3 client to download the file from the specified bucket and write it to the specified file path.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

Prompt template for Anthropic Claude """ Human: Write a function in {{programming language}} to {{task or purpose}} Assistant:"""
User prompt: Human: Write some python code that quickly approximates the square root of a number. Assistant:
Output: Here is a simple function to approximate the square root of a number using Newton's method: ```python def sqrt(n, iterations=10): x = n for i in range(iterations): x = (x + n/x) / 2 return x ``` To use it: ```python print(sqrt(25)) # 5.0 print(sqrt(15, iterations=20)) # 3.872983346207417 ``` This repeatedly takes the average of x and n/x to converge on the square root. More iterations produce a more accurate approximation. The advantage of this approach is it converges very quickly, so even just 10 iterations produces a decent estimate.

(Fonte do prompt: AWS, modelo usado: Claude da Anthropic)

Matemática

A entrada descreve um problema que requer raciocínio matemático em algum nível, que pode ser numérico, lógico, geométrico ou outro. Para esse problema, é útil pedir ao modelo que resolva o problema de forma fragmentada, adicionando frases às instruções, como Let’s think step by step ou Think step by step to come up with the right answer.

Prompt template for Titan """{{Math problem description}} Let’s think step by step."""
User prompt: A triangle has two angles of 70 degrees and 50 degrees. What is the third angle in degrees? Let’s think step by step.
Output: The sum of the three angles in a triangle is always 180 degrees. The two given angles are 70 degrees and 50 degrees. 70 + 50 = 120. Therefore, the third angle is 180 - 120 = 60 degrees. The third angle is 60 degrees.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

O exemplo a seguir também demonstra como pedir ao modelo que pense passo a passo:

Prompt template for Titan """{{Math problem description}} Think step by step to come up with the right answer."""
User prompt: Lucy has 12 colorful marbles, and she wants to share them equally with her 4 friends. How many marbles will each friend receive? Think step-by-step to come up with the right answer.
Output: 1. Lucy has 12 marbles and wants to share them equally with her 4 friends, then each friend will receive 12 / 4 = 3 marbles. 2. Therefore, each friend will receive 3 marbles.

(Fonte do prompt: AWS, modelo usado: Jurassic-2 Ultra da AI21 Labs v1)

Raciocínio/pensamento lógico

Em tarefas de raciocínio complexas ou problemas que exijam raciocínio lógico, é possível pedir ao modelo que faça deduções lógicas e explique suas respostas.

Prompt template for Titan and AI21 Labs Jurassic: """Question: {{question}} Please output the answer and then explain your answer:"""
User prompt: Question: Which word is the odd one out? A. accomplished B. good C. horrible D. outstanding Please output the answer and then explain your answer:
Output: The odd one out is C. horrible. The other three words describe positive qualities or achievements, whereas horrible is an negative adjective.

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text)

Veja outro exemplo usando o modelo Claude da Anthropic:

Prompt template for Anthropic Claude: """ Human: {{question}} Please provide the answer and show the reasoning. Assistant:"""
User prompt Human: The barber is the "one who shaves all those, and those only, who do not shave themselves". Does the barber shave himself? Why is this a paradox? Please provide the answer and show the reasoning. Assistant:
Output: This is an example of a paradox known as the Barber paradox. Here is an explanation: - The statement defines a barber who shaves all and only those people who do not shave themselves. - So if the barber shaves himself, he must shave all those who do not shave themselves, which includes himself. But the barber cannot shave himself as per the initial statement. - If the barber does not shave himself, then he would fit into the group of people who do not shave themselves, and so he must shave himself, according to the statement. But again this contradicts the initial statement. - So the barber cannot shave himself and also cannot not shave himself. This is self-contradictory based on the initial statement about the barber. - Therefore, the statement leads to a paradox where both possibilities (the barber shaving himself or not shaving himself) result in a contradiction. In summary, the paradox arises because the definition of the barber's behavior is self-contradictory when applied to the barber himself. This makes it impossible to determine if the barber shaves himself or not based on the given statement alone.

(Fonte do prompt: https://en.wikipedia.org/wiki/Barber_paradox, modelo usado: Claude da Anthropic)

Extração de entidades

Para extração de entidade de uma pergunta de entrada fornecida. Extraia as entidades do texto gerado e coloque-as em tags XML para processamento adicional.

Prompt template for Titan """You are an expert entity extractor from provided input question. You are responsible for extracting following entities: {{ list of entities}} Please follow below instructions while extracting the entity A, and reply in <entityA> </entityA> XML Tags: {{ entity A extraction instructions}} Please follow below instructions while extracting the entity B, and reply in <entityB> </entityB> XML Tags: {{ entity B extraction instructions}} Below are some examples: {{ some few shot examples showing model extracting entities from give input }}

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text G1 - Premier)

Exemplo:

User: You are an expert entity extractor who extracts entities from provided input question. You are responsible for extracting following entities: name, location Please follow below instructions while extracting the Name, and reply in <name></name> XML Tags: - These entities include a specific name of a person, animal or a thing - Please extract only specific name name entities mentioned in the input query - DO NOT extract the general mention of name by terms of "name", "boy", "girl", "animal name", etc. Please follow below instructions while extracting the location, and reply in <location></location> XML Tags: - These entities include a specific location of a place, city, country or a town - Please extract only specific name location entities mentioned in the input query - DO NOT extract the general mention of location by terms of "location", "city", "country", "town", etc. If no name or location is found, please return the same input string as is. Below are some examples: input: How was Sarah's birthday party in Seattle, WA? output: How was <name>Sarah's</name> birthday party in <location>Seattle, WA</location>? input: Why did Joe's father go to the city? output: Why did <name>Joe's</name> father go to the city? input: What is the zipcode of Manhattan, New york city? output: What is the zipcode of <location>Manhattan,New york city<location>? input: Who is the mayor of San Francisco? Bot:

Raciocínio da cadeia de pensamento

Forneça uma análise passo a passo sobre como a resposta foi obtida. Verifique os fatos e valide como o modelo produziu uma resposta.

Prompt template for Titan """ {{Question}} {{ Instructions to Follow }} Think Step by Step and walk me through your thinking """

(Fonte do prompt: AWS, modelo usado: Amazon Titan Text G1 - Premier)

Exemplo:

User: If Jeff had 100 dollars, and he gave $20 to Sarah, and bought lottery tickets with another $20. With the lottery tickets he bought he won 35 dollars. Jeff then went to buy his lunch and spend 40 dollars in lunch. Lastly he made a donation to charity for $20. Stephen met with Jeff and wanted to lend some money from him for his taxi. How much maximum money can Jeff give to Stephen, given that he needs to save $10 for his ride back home?. Please do not answer immediately, think step by step and show me your thinking. Bot: