 Ver em ai.google.dev
|
 Executar no Google Colab
|
 Consulte o código-fonte no GitHub
|
Neste notebook, você vai aprender a usar a API PaLM, que dá acesso aos modelos de linguagem grandes mais recentes do Google. Aqui, você vai aprender a usar os recursos de geração de texto da API PaLM.
Configuração
Primeiro, faça o download e instale a biblioteca Python da API PaLM.
pip install -q google-generativeai
import pprint
import google.generativeai as palm
Obter uma chave de API
Para começar, crie uma chave de API.
palm.configure(api_key='YOUR_API_KEY')
Geração de texto
Use a função palm.list_models para encontrar modelos disponíveis:
models = [m for m in palm.list_models() if 'generateText' in m.supported_generation_methods]
model = models[0].name
print(model)
models/text-bison-001
Use o método palm.generate_text para gerar texto:
prompt = """
You are an expert at solving word problems.
Solve the following problem:
I have three houses, each with three cats.
each cat owns 4 mittens, and a hat. Each mitten was
knit from 7m of yarn, each hat from 4m.
How much yarn was needed to make all the items?
Think about it step by step, and show your work.
"""
completion = palm.generate_text(
model=model,
prompt=prompt,
temperature=0,
# The maximum length of the response
max_output_tokens=800,
)
print(completion.result)
There are 3 houses * 3 cats / house = 9 cats. So, 9 cats * 4 mittens / cat = 36 mittens were made. Also, 9 cats * 1 hat / cat = 9 hats were made. So, 36 mittens * 7m / mitten = 252m of yarn was used for the mittens. Also, 9 hats * 4m / hat = 36m of yarn was used for the hats. In total, 252m + 36m = 288m of yarn was used. Thus, the answer is 288.
Mais opções
A função palm.generate_text tem alguns outros argumentos que valem a pena mencionar.
Sequências de paradas
Use o argumento stop_sequences para interromper a geração antecipadamente.
Por exemplo, os LLMs muitas vezes cometem erros na aritmética. Você poderia pedir ao modelo para "usar uma calculadora" colocando equações em uma tag <calc>.
Faça com que o modelo pare na tag de fechamento para que você possa editar o comando:
calc_prompt = f"""
Please solve the following problem.
{prompt}
----------------
Important: Use the calculator for each step.
Don't do the arithmetic in your head.
To use the calculator wrap an equation in <calc> tags like this:
<calc> 3 cats * 2 hats/cat </calc> = 6
----------------
"""
equation=None
while equation is None:
completion = palm.generate_text(
model=model,
prompt=calc_prompt,
stop_sequences=['</calc>'],
# The maximum length of the response
max_output_tokens=800,
)
try:
response, equation = completion.result.split('<calc>', maxsplit=1)
except Exception:
continue
print(response)
Chain-of-thought: There are three houses, and each house has three cats, so there are 3 houses * 3 cats / house = 9 cats. Each cat has 4 mittens, so the cats need 9 cats * 4 mittens / cat = 36 mittens. Each mitten takes 7m of yarn, so 36 mittens * 7m / mitten = 252m of yarn. Each cat has a hat, and each hat takes 4m of yarn, so 9 cats * 4m / cat = 36m of yarn. So, in total, 36m + 252m = 288m of yarn were needed. The answer should be
print(equation)
9 cats * 4 mittens / cat
A partir daí, é possível calcular o resultado e elaborar um novo comando para o modelo continuar. Para uma implementação completa, consulte o exemplo de Calculadora de texto.
Candidatos
Normalmente, há um grau de aleatoriedade no texto produzido pelos LLMs. Saiba o motivo na introdução do LLM (link em inglês). Isso significa que, ao chamar a API mais de uma vez com a mesma entrada, você pode receber respostas diferentes. Esse recurso pode ser usado a seu favor para obter respostas de modelo alternativas.
O argumento temperature controla a variância das respostas. O objeto palm.Model fornece o valor padrão para temperature e outros parâmetros.
models[0]
Model(name='models/text-bison-001', base_model_id='', version='001', display_name='Text Bison', description='Model targeted for text generation.', input_token_limit=8196, output_token_limit=1024, supported_generation_methods=['generateText'], temperature=0.7, top_p=0.95, top_k=40)
O argumento candidate_count controla o número de respostas retornadas:
completion = palm.generate_text(
model=model,
prompt=prompt,
# The number of candidates to return
candidate_count=8,
# Set the temperature to 1.0 for more variety of responses.
temperature=1.0,
max_output_tokens=800,
)
print(completion.result)
In each house there are 3 cats * 4 mittens / cat = 12 mittens. In total there are 3 houses * 12 mittens / house = 36 mittens. In total there are 36 mittens * 7m / mitten = 252m of yarn for the mittens. In total there are 3 houses * 3 cats / house * 1 hat / cat = 9 hats. In total there are 9 hats * 4m / hat = 36m of yarn for the hats. In total there are 36m yarn for the hats + 252m yarn for the mittens = 288m of yarn. The answer: 288.
Quando você solicita vários candidatos, o atributo Completion.result ainda contém apenas o primeiro. O atributo Completion.candidates contém todos eles:
import pprint
pprint.pprint(completion.candidates)
[{'output': 'In each house there are 3 cats * 4 mittens / cat = 12 mittens. In '
'total there are 3 houses * 12 mittens / house = 36 mittens. In '
'total there are 36 mittens * 7m / mitten = 252m of yarn for the '
'mittens. In total there are 3 houses * 3 cats / house * 1 hat / '
'cat = 9 hats. In total there are 9 hats * 4m / hat = 36m of yarn '
'for the hats. In total there are 36m yarn for the hats + 252m '
'yarn for the mittens = 288m of yarn.\n'
'The answer: 288.',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'Each house has 3 cats, so each house needs 3 * 4 = 12 mittens. '
"With three houses, that's 3 * 12 = 36 mittens. And each house "
'needs 3 * 1 = 3 hats. So in total, we need 3 hats + 36 mittens = '
'39 items. Each mitten needs 7 meters of yarn, so 39 mittens need '
'39 * 7 = 273 meters of yarn. Each hat needs 4 meters of yarn, and '
"we need 3 hats, so that's 4 * 3 = 12 meters of yarn. So in total, "
'we needed 12 + 273 = 285 meters of yarn.\n'
'Thus, the answer is 285.',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'There are 3 houses * 3 cats / house = 9 cats. There are 9 cats * '
'4 mittens / cat = 36 mittens. There are 9 cats * 1 hat / cat = 9 '
'hats. The total amount of yarn for the mittens is 36 mittens * 7m '
'/ mitten = 252m. The total amount of yarn for the hats is 9 hats '
'* 4m / hat = 36m. The total amount of yarn is 252m + 36m = 288m.\n'
'Thus, the answer is 288.',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'There are 3 houses * 3 cats / house = 9 cats. Each cat has 4 '
'mittens + 1 hat = 5 items. So the total number of items is 9 cats '
'* 5 items / cat = 45 items. Thus, 45 items * 7m / item = 315m of '
'yarn was needed.\n'
'Thus, the answer is 315.',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'Chain-of-thought:\n'
'There are 3 houses * 3 cats / house = 9 cats.\n'
'The cats need 9 cats * 4 mittens / cat = 36 mittens.\n'
'The cats need 9 cats * 1 hat / cat = 9 hats.\n'
'The mittens need 36 mittens * 7m / mitten = 252m of yarn.\n'
'The hats need 9 hats * 4m / hat = 36m of yarn.\n'
'Therefore, the total amount of yarn needed is 252m + 36m = 288m.\n'
'\n'
'The answer should be 288',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'First find the total number of cats: 3 houses * 3 cats / house = '
'9 cats. Then multiply that number by the number of mittens per '
'cat to find the total number of mittens: 9 cats * 4 mittens / cat '
'= 36 mittens. Then multiply that number by the number of meters '
'of yarn per mitten to find the total amount of yarn used for '
'mittens: 36 mittens * 7 meters / mitten = 252 meters. Then do the '
'same thing for hats: 9 cats * 1 hat / cat = 9 hats. Then multiply '
'that number by the number of meters of yarn per hat to find the '
'total amount of yarn used for hats: 9 hats * 4 meters / hat = 36 '
'meters. Then add the amount of yarn used for mittens and hats to '
'find the total amount of yarn used: 36 meters + 252 meters = 288 '
'meters.\n'
'Thus, the answer is 288.',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'The total number of cats is 3 houses * 3 cats / house = 9 cats. '
'So, the total number of mittens is 9 cats * 4 mittens / cat = 36 '
'mittens. The total number of hats is 9 cats * 1 hat / cat = 9 '
'hats. The total length of yarn needed to make the mittens is 36 '
'mittens * 7 m / mitten = 252 m. The total length of yarn needed '
'to make the hats is 9 hats * 4 m / hat = 36 m. So, the total '
'length of yarn needed is 252 m + 36 m = 288 m.\n'
'\n'
'The answer: 288',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]},
{'output': 'There are 3 houses with 3 cats each, so 3 * 3 = 9 cats. Each cat '
'has 4 mittens and a hat, so 9 cats * 4 mittens / cat + 9 cats * 1 '
'hat / cat = 36 mittens and 9 hats. Each mitten takes 7m of yarn '
'and each hat takes 4m of yarn, so the total yarn needed is 36 '
'mittens * 7m / mitten + 9 hats * 4m / hat = 252m + 36m = 288m.\n'
'The answer: 288.',
'safety_ratings': [{'category': <HarmCategory.HARM_CATEGORY_DEROGATORY: 1>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_TOXICITY: 2>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_VIOLENCE: 3>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_SEXUAL: 4>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_MEDICAL: 5>,
'probability': <HarmProbability.NEGLIGIBLE: 1>},
{'category': <HarmCategory.HARM_CATEGORY_DANGEROUS: 6>,
'probability': <HarmProbability.NEGLIGIBLE: 1>}]}]
Então, como você sabe a resposta para esse problema, é fácil verificar a taxa de resolução:
import numpy as np
np.mean(['288' in c['output'] for c in completion.candidates])
0.75