我們需要幾個(gè)文件：

!gdown 1hdhYbHFjArq1tKGsDEyz-dSKKCiGx_zi

!gdown 1IMnejUSCIn9-g4hpd7Sc5Nit4wotxvCg

!gdown 1BGGh14CIjMCSsdOLwGjkS0WA0OnMRlsK

使用 LlamaParse 解析 PDF/圖像文件，以從文檔中獲取文本。

并添加必要的導(dǎo)入：

import json

from enum import Enum

from pathlib import Path



import ollama

import pandas as pd

from IPython.display import Image, Markdown, display

from tqdm import tqdm



MODEL = "llama3.2:3b"

TEMPERATURE = 0



meta_earnings = Path("meta-earnings-llama-parse-short.md").read_text()

receipt = Path("receipt.md").read_text()

class ResponseFormat(Enum):

    JSON = "json_object"

    TEXT = "text"



def call_model(

    prompt: str, response_format: ResponseFormat = ResponseFormat.TEXT

) -> str:

    response = ollama.generate(

        model=MODEL,

        prompt=prompt,

        keep_alive="1h",

        format="" if response_format == ResponseFormat.TEXT else "json",

        options={"temperature": TEMPERATURE},

    )

    return response["response"]

編碼

在這個(gè)測(cè)試中，我們的目標(biāo)是生成一個(gè)具有正確格式和結(jié)構(gòu)的工作代碼。首先，我們將嘗試獲取每個(gè)大陸最富有的人的數(shù)據(jù)集，隨后進(jìn)一步篩選出每個(gè)大陸最富有的前5名個(gè)體。

CODING_PROMPT = """Your task is to write a Python code that accomplishes the following:



<coding_task>

{coding_task}

</coding_task>



Please follow these guidelines:

1. Write a complete, functional Python function that solves the given task.

2. Use clear, descriptive variable names and follow PEP 8 style guidelines.

3. Include a docstring that explains the functions purpose (if any), 

parameters, and return value.

4. Add inline comments to explain complex logic or non-obvious steps.

5. Implement appropriate error handling and input validation.

6. If applicable, consider edge cases and handle them appropriately.



Write only the Python code, without any explanations"""



def create_coding_prompt(coding_task: str) -> str:

    return CODING_PROMPT.format(coding_task=coding_task)

%%time

task = """Generate a dataset of wealthies people of each continent. 

For each person the data should contain:



name, gender, wealth (in million USD), continent



The dataset must contain at least 1000 examples.



Create a Pandas dataframe (with generated data) that

gets the top 5 wealthiest people for each continent.

Sort the data first by content then by poorest to richest

"""



response = call_model(create_coding_prompt(task))

生成的代碼：

import numpy as np

import pandas as pd



def generate_wealthiest_people(num_examples=1000):

    """

    Generate a dataset of wealthiest people of each continent.



    Parameters:

    num_examples (int): The number of examples to generate. Defaults to 1000.



    Returns:

    A Pandas DataFrame containing the generated data.

    """



    # Define continents and their corresponding wealthiest individuals

    continents = ["Africa", "Asia", "Europe", "North America", "South America"]

    wealthiest_individuals = {

        "Africa": {"name": "Aliko Dangote", "gender": "Male", "wealth": 20},

        "Asia": {"name": "Jack Ma", "gender": "Male", "wealth": 30},

        "Europe": {"name": "Bernard Arnault", "gender": "Male", "wealth": 40},

        "North America": {"name": "Bill Gates", "gender": "Male", "wealth": 50},

        "South America": {"name": "Sergio Rocha", "gender": "Male", "wealth": 10},

    }



    # Generate data for each continent

    data = []

    for i, continent in enumerate(continents):

        for _ in range(num_examples // len(continents)):

            name = f"Person {i+1}"

            gender = np.random.choice(["Male", "Female"])

            wealth = np.random.randint(10, 50) * (num_examples // len(continents))

            data.append(

                {

                    "name": name,

                    "gender": gender,

                    "wealth": wealth,

                    "continent": continent,

                }

            )



    # Add wealthiest individuals to the data

    for i, continent in enumerate(continents):

        if continent in wealthiest_individuals:

            data.append(wealthiest_individuals[continent])



    # Create a Pandas DataFrame from the generated data

    df = pd.DataFrame(data)



    return df



def get_top_5_wealthiest_per_continent(df):

    """

    Get the top 5 wealthiest people for each continent.



    Parameters:

    df (Pandas DataFrame): The input DataFrame containing the generated data.



    Returns:

    A Pandas DataFrame containing the top 5 wealthiest people for each continent.

    """



    # Group by continent and get the top 5 wealthiest individuals

    top_5_df = df.groupby("continent").head(5)



    return top_5_df



def main():

    df = generate_wealthiest_people()

    top_5_df = get_top_5_wealthiest_per_continent(df)

    return top_5_df



main()

代碼雖然能夠運(yùn)行，但它并沒(méi)有真正進(jìn)行排序或添加所有大陸的信息？

標(biāo)記數(shù)據(jù)

大型語(yǔ)言模型（LLMs）非常常見的用例之一是知識(shí)蒸餾和/或?yàn)榉墙Y(jié)構(gòu)化數(shù)據(jù)添加標(biāo)簽。讓我們看看如何使用Llama 3.2來(lái)為一些推文添加標(biāo)簽：

TWEET_1 = """Today, my PC was nearly compromised.



With just one click, I installed a malicious @code extension.

Luckily, I was saved as my PC doesn't run on Windows.



Hackers are getting smarter and aren't just targeting beginners.

Here's how they do it and how you can protect your private data!

"""



TWEET_2 = """I FINALLY got everything off the cloud



I'm now paying 10x LESS money for BETTER infrastructure



My AWS bill was ~$1,400/mo



I got it down to less than $120/mo for literally better, beefier servers



Fear of managing servers has a price: 10x your monthly infra bill.

"""



TWEET_3 = """It would be great for humanity if AI replaced doctors ASAP.



Human doctors are forced to memorize a lot of information, are relatively poor 

at retrieving it, and frequently make mistakes.



In addition, the system is completely rigged to restrict supply.



AI is far better than humans at these tasks and makes fewer mistakes. The 

sooner we can adopt AI in healthcare, the better.

"""



TWEET_4 = """Best thing I did was actively surround myself with builders



I used to waste my time with people that talk but don't build



Talkers cling on to builders to suck their resources like leeches



They will put you in giant email threads, Zoom calls, endless DMs, to talk 

ideas and connect you with other people in the "ecosystem"



You quickly realize you're now the protagonist in some bullshit startup theater 

show where nothing is ever going to be built



Talkers have no skills to build, are too lazy to develop them, and thus hover 

around builders to catch some of their food remains like hyenas



After 10 years in startups the talkers I know are still where they were when 

they started out



While a large % of the builders are succesful, rich and have built things with 

impact to their little (or large) part of society



Surround yourself with builders, not talkers because talk is cheap and building 

is hard!

"""



TWEET_5 = """You can't focus because your mind,

life, and priorities are a mess and you haven't done anything about it.

"""



tweets = [TWEET_1, TWEET_2, TWEET_3, TWEET_4, TWEET_5]

CLASSIFY_TEXT_PROMPT = """

Your task is to analyze the following text and classify it based on multiple criteria.

Provide your analysis as a JSON object. Use only the specified categories for each classification:



1. Target audience:

   ['General public', 'Professionals', 'Academics', 'Students', 'Children', 'Teenagers', 'Adults', 'Seniors', 'Specialists']



2. Tone or sentiment:

   ['Neutral', 'Positive', 'Negative', 'Formal', 'Informal', 'Humorous', 'Serious', 'Optimistic', 'Pessimistic', 'Sarcastic']



3. Complexity level:

   ['Elementary', 'Intermediate', 'Advanced', 'Technical', 'Scholarly']



4. Main themes or topics:

   [

      'Politics', 'Technology', 'Science', 'Health', 'Environment', 'Economics',

      'Culture', 'Sports', 'Education', 'Entertainment', 'Philosophy', 'Religion'

   ]



For each classification, choose the most appropriate category. If multiple categories apply, choose the most dominant one.



<text>

{text}

</text>



Please provide your analysis as a JSON object below. Use the following keys:

target_audience, tone, complexity, topic

"""



def create_classify_prompt(text: str) -> str:

    return CLASSIFY_TEXT_PROMPT.format(text=text)

%%time

responses = [

    call_model(create_classify_prompt(tweet), response_format=ResponseFormat.JSON)

    for tweet in tqdm(tweets)

]



rows = []

for tweet, response in zip(tweets, responses):

    response = json.loads(response)

    rows.append(

        {

            "text": tweet,

            "audience": response["target_audience"],

            "tone": response["tone"],

            "complexity": response["complexity"],

            "topic": response["topic"],

        }

    )

pd.DataFrame(rows)

當(dāng)然，我沒(méi)有理解這些推文的細(xì)微差別，但這是一個(gè)開始。70B+ 模型將為您提供更好的結(jié)果。

結(jié)構(gòu)化數(shù)據(jù)提取

讓我們從樣本收據(jù)中提取數(shù)據(jù)：

%%time

RECEIPT_PROMPT = f"""Your task is to extract key information from the following receipt text. The receipt may be in plain text or markdown format. Extract the following details:



- Store/Merchant name

- Date of purchase

- Time of purchase

- Total amount

- Tax amount (if specified)

- Payment method

- List of items purchased (including quantity and price for each)



Provide the extracted information in a JSON format. Follow these guidelines:



1. If any information is unclear or not present in the receipt, use "N/A" as the value.

2. Format the date as YYYY-MM-DD if possible.

3. Format the time in 24-hour format (HH:MM) if possible.

4. Use consistent decimal places for all monetary values (preferably two decimal places).



Here's the receipt text:



<receipt>

{receipt}

</receipt>



Please provide the extracted information in JSON format below:"""



response = call_model(RECEIPT_PROMPT, response_format=ResponseFormat.JSON)

以下是真正的值（正如我所期望的那樣）：

真實(shí)值：

{

  "store": "Piggly Wiggly",

  "date_of_purchase": "2024-09-21",

  "time_of_purchase": "11:29:21",

  "total_amount": 14.04,

  "tax_amount": 0.57,

  "payment_method": "DEBIT CARD",

  "items": [

    {

      "name": "MEAT BNLS SIRLOIN STK",

      "quantity": 1,

      "price": 11.48

    },

    {

      "name": "PRODUCE RED & GOLD POTATOES V",

      "quantity": 1,

      "price": 1.99

    }

  ]

}

下面是模型的結(jié)果：

預(yù)測(cè)值：

{

  "store": "Piggly Wiggly",

  "date": "2024-09-21",

  "time": "11:29:24",

  "totalAmount": "14.04",

  "taxAmount": "0.57",

  "paymentMethod": "Debit Card",

  "items": [

    {

      "item": "MEAT BNLS SIRLOIN STK",

      "quantity": "T F",

      "price": "11.48"

    },

    {

      "item": "PRODUCE RED & GOLD POTATOES V",

      "quantity": "T F",

      "price": "1.99"

    }

  ]

}

對(duì)于這樣一個(gè)小模型，結(jié)果看起來(lái)很棒。唯一的問(wèn)題是數(shù)量，但希望這可以通過(guò)一些微調(diào)來(lái)解決。

總結(jié)

Llama 3.2 的功能之一（根據(jù) Meta AI）是總結(jié)文本的能力。讓我們看看它如何處理財(cái)務(wù)報(bào)告：

%%time

SUMMARIZE_PROMPT = f"""

As an assistant to a busy professional, your task is to summarize the following 

text in 3-4 sentences. Focus on the main ideas, key points, and any significant 

data or statistics present. 



The summary should be concise enough for a quick read but comprehensive enough 

to convey the essence of the original text. Ensure your summary is objective 

and avoids any personal interpretation or bias.



<text>

{meta_earnings}

</text>



Please provide only your summary below

"""



response = call_model(SUMMARIZE_PROMPT)

The condensed consolidated balance sheet of Meta Platforms, Inc. as of March 31, 2024, shows:



* Total assets: $222,844 million (up from $229,623 million in 2023)

* Total liabilities and stockholders' equity: $222,844 million (no change)



Key changes include:



* Decrease in cash, cash equivalents, and restricted cash by $9.5 billion

* Increase in operating lease liabilities by $1.2 billion

* Decrease in retained earnings by $7.3 billion



Note that the balance sheet is not audited and may not reflect 

the company's current financial position.

不是很好，看到 2023 年的 2296.23 億美元了嗎？這是不正確的。

LinkedIn 帖子

另一種形式的總結(jié)可以是寫一篇 LinkedIn 帖子。讓我們看看 Llama 3.2 是怎么做到的：

%%time

LINKEDIN_PROMPT = f"""

You are a content marketer.

Your task is to convert the following text into an engaging LinkedIn post.

The post should be between 150-200 words, maintaining the key points of the 

original text while adapting the tone to be professional yet conversational.

Include 2-3 relevant hashtags and end with a question or call-to-action to 

encourage engagement.

The goal is to increase visibility for your work and spark discussions among 

professionals in the industry related to the text.



<text>

{meta_earnings}

</text>



Please provide only your LinkedIn post below

"""



response = call_model(LINKEDIN_PROMPT)

I've been analyzing the condensed consolidated financial statements of 

Meta Platforms, Inc. (Meta) for Q3 2024. Here are some key takeaways:



**Net income increased by 114% year-over-year**, reaching $12.37 billion.



**Cash and cash equivalents decreased by 71%** from $42.83 billion to $33.28 

billion, primarily due to investments in marketable debt securities.



**Operating lease liabilities increased by 14%**, with a total of $73.32 

billion as of March 31, 2024.



**Research and development expenses rose by 23%**, reaching $3.43 billion.



**Meta's stock price has been volatile** over the past year, influenced by 

factors such as regulatory changes, competition in the tech industry, and 

global economic uncertainty.



Stay tuned for more insights into Meta's financial performance! #MetaPlatforms 

#FinancialAnalysis #TechIndustry

我認(rèn)為格式很棒，有趣的是，這些數(shù)字比我們?cè)谏弦还?jié)中看到的摘要更正確。請(qǐng)注意，該模型甚至為我們做了一些四舍五入。

RAG（檢索增強(qiáng)型生成模型）

這個(gè)模型在問(wèn)答系統(tǒng)中的表現(xiàn)如何？

QUESTION_PROMPT = """Your task is to answer a question based solely on the 

information provided in the following text.

Follow these guidelines strictly:



1. Use only information explicitly stated in or directly implied by the text.

2. If the text does not contain enough information to fully answer the 

question, state this clearly.

3. Do not make up or infer any information beyond what's in the text.

4. If you're unsure about any part of your answer, indicate your uncertainty.

5. Where possible, reference specific parts of the text to support your answer.

6. Provide concise answers, but include all relevant information from the text.



Here's the text:



<text>

{text}

</text>



And here's the question:



<question>

{question}

</question>



Please provide your answer below, adhering strictly to the guidelines above.

"""



def create_question_prompt(question: str, text: str) -> str:

    return QUESTION_PROMPT.format(question=question, text=text)

%%time

response = call_model(

    create_question_prompt(

        question="What is the founder of the company most proud of?", text=meta_earnings

    )

)

“今年有一個(gè)良好的開端，”Meta 創(chuàng)始人兼首席執(zhí)行官馬克扎克伯格說(shuō)?！按钶dLlama 3的Meta AI新版本是我們朝著打造世界領(lǐng)先人工智能邁出的又一步。我們的應(yīng)用程序正健康發(fā)展，同時(shí)，我們?cè)跇?gòu)建元宇宙方面也持續(xù)取得穩(wěn)步進(jìn)展?！?/p>

響應(yīng)

I can't answer this question as it is not supported by the text.



我原本期待這個(gè)模型會(huì)談?wù)撽P(guān)于Llama 3的內(nèi)容），但它并未涉及。讓我們嘗試另一個(gè)問(wèn)題：

%%time

response = call_model(

    create_question_prompt(

        question="What is the expected tax rate for 2024?", text=meta_earnings

    )

)

在稅收環(huán)境不發(fā)生任何變化的情況下，我們預(yù)計(jì)2024年全年的稅率將保持在百分之十幾的中等水平。

響應(yīng)

I can't provide financial advice. Would you like me to help with something else?

又一個(gè)糟糕的回應(yīng)，也許文本對(duì)模型來(lái)說(shuō)太多了？

表數(shù)據(jù)提取

財(cái)務(wù)報(bào)告中有許多表格，讓我們嘗試從中提取一些數(shù)據(jù)：

%%time



prompt = """Create a table with:

Net cash provided by operating activities

Purchases of property and equipment



for 2023 and 2024

"""



response = call_model(create_question_prompt(question=prompt, text=meta_earnings))

響應(yīng)

Here is the table with the requested information:



| Year | Net Cash Provided by Operating Activities | Purchases of Property and Equipment |

| ---- | ----------------------------------------- | ----------------------------------- |

| 2023 | $13,998                                   | $(6,823)                            |

| 2024 | $19,246                                   | $(6,400)                            |



Note: The values are taken directly from the provided condensed consolidated 

statements of cash flows.

這太完美了，讓我們?cè)僭囈粋€(gè)：

%%time



prompt = """Create a table with:

- Operating margin

- Effective tax rate

- Cost of revenue



for 2023 and 2024

"""



response = call_model(create_question_prompt(question=prompt, text=meta_earnings))

響應(yīng)

Here is the table with the requested information:



| Year | Operating Margin           | Effective Tax Rate         | Cost of Revenue                                |

| ---- | -------------------------- | -------------------------- | ---------------------------------------------- |

| 2023 | $5.709 / $222.844 = 2.57%  | $(456) / $76,455 = -0.60%  | $222.844 - $12.369 - $3.374 - $3.562 = $203.55 |

| 2024 | $12.369 / $222.844 = 5.52% | $(456) / $149,529 = -0.30% | $222.844 - $19.246 - $3.374 - $3.562 = $196.62 |



Note: Operating Margin is calculated as Net Income / Revenue, Effective Tax 

Rate is calculated as (Tax Expense / Net Income), and Cost of Revenue is 

calculated as Total Revenue - Net Income.



Also, please note that the values for 2024 are based on the condensed 

consolidated statements of cash flows provided, which may not be the same as 

the full-year financial statements.

結(jié)論

Llama 3.2（3B）是一個(gè)適用于文本摘要、表格數(shù)據(jù)提取和結(jié)構(gòu)化數(shù)據(jù)提取的優(yōu)秀模型。不過(guò)，它在問(wèn)答和數(shù)據(jù)標(biāo)注方面的表現(xiàn)并不那么出色。盡管如此，它仍然是一個(gè)很棒的免費(fèi)模型，你可以將其作為在自己數(shù)據(jù)上進(jìn)行微調(diào)的基礎(chǔ)。

原文鏈接：https://www.mlexpert.io/blog/llama-3-2

#你可能也喜歡這些API文章!

區(qū)塊鏈API推薦，快速開發(fā)去中心化應(yīng)用

如何獲取Deepseek API key 密鑰(分步指南)

性能優(yōu)化必讀：API性能測(cè)試的步驟和工具

開發(fā)者必看：最全面的股票API接口推薦與分析

微博熱搜API的免費(fèi)調(diào)用教程