以下是本次课程的正文：

在当今大模型快速发展的时代，与AI相关的新技术、新算法以及新产品层出不穷，几乎每天都能看到新的突破和应用。然而，很多前沿技术和动态的介绍通常以英文形式呈现，这给科研人员和技术工程师带来了一定的挑战。对于从事一线研究和开发的人员来说，掌握最新动态至关重要。那么，如何高效获取这些信息呢？一般来说，主要有两个常见渠道：

第二种 是直接阅读英文原文资料，比如论文、技术文档或博客。然而，这种方式对阅读者的英文水平要求较高。对于一些人来说，面对长篇英文文档可能感到吃力，尤其是在快速获取关键信息的场景中，难免会有理解障碍。

25.1 整体业务流程

25.2 核心模块

25.2.1 抽取文本

[    [        {            "page_num": 1,            "text": "OpenAI o1 System Card",            "top": 113,            "bottom": 130,            "x0": 210.437,            "x1": 384.83416662        },        {            "page_num": 1,            "text": "OpenAI",            "top": 160,            "bottom": 170,            "x0": 280.346,            "x1": 314.9311659        },        {            "page_num": 1,            "text": "December 5, 2024",            "top": 193,            "bottom": 205,            "x0": 252.412,            "x1": 343.26793344        },        {            "page_num": 1,            "text": "1 Introduction",            "top": 248,            "bottom": 263,            "x0": 70.86600000000001,            "x1": 181.25427052000003        },        {            "page_num": 1,            "text"




    
: "The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.",            "top": 284,            "bottom": 444,            "x0": 70.476,            "x1": 525.9228737591997        },        {            "page_num": 1,            "text": "2 Model data and training",            "top": 478,            "bottom": 492,            "x0": 70.866,            "x1": 264.6329503        },        {            "page_num": 1,            "text": "The o1 large language model family is trained with reinforcement learning to perform complex reasoning. o1 thinks before it answers—it can produce a long chain of thought before responding to the user. OpenAI o1 is the next model in this series (previously OpenAI o1-preview), while OpenAI o1-mini is a faster version of this model that is particularly effective at coding. Through training, the models learn to refine their thinking process, try different strategies, and recognize their mistakes. Reasoning allows o1 models to follow specific guidelines and model policies we’ve set, helping them act in line with our safety expectations. This means they are better at providing helpful answers and resisting attempts to bypass safety rules, to avoid producing unsafe or inappropriate content.",            "top": 514,            "bottom": 633,            "x0": 70.476,            "x1": 524.7064330704997        },        {            "page_num": 1,            "text": "The two models were pre-trained on diverse datasets, including a mix of publicly available data, proprietary data accessed through partnerships, and custom datasets developed in-house, which collectively contribute to the models’ robust reasoning and conversational capabilities.",            "top": 647,            "bottom": 685,            "x0": 70.476,            "x1": 525.9197562634197        },        {            "page_num": 1,            "text": "Select Public Data: Both models were trained on a variety of publicly available datasets, including web data and open-source datasets. Key components include reasoning data and scientific literature. This ensures that the models are well-versed in both general knowledge and technical topics, enhancing their ability to perform complex reasoning tasks.",            "top": 712,            "bottom": 762,            "x0": 70.866,            "x1": 525.9128251106        },        {            "page_num": 1,            "text": "1",            "top": 790,            "bottom": 801,            "x0": 294.926,            "x1": 300.35000451999997        }    ]]

{        "page_num": 1,   # 页码        "text": "OpenAI o1 System Card",  # 段的文本        "top": 113,  # 这一段最上面距离页面最上端的像素个数        "bottom": 130,  # 这一段最下面距离页面最上端的像素个数        "x0": 210.437,  # 这一段最左边距离页面最左端的像素个数        "x1": 384.83416662  # 这一段最右边边距离页面最左端的像素个数}

import pdfplumber
def get_paragraph_lines_info(paragraph_lines: list[dict]) -> dict:    """    提取段落的相关信息：合并文本，获取段落边界框（top, bottom, x0, x1）。
    :param paragraph_lines:        [                {                    "text": "Given the lack of standardized datasets specific to chip design, we adopted the ChipNeMo [90]",                    "top": 194,                    "bottom": 205,                    "x0": 72.0,                    "x1": 540.0045790943997                },                ...        ]    :return:        {            "text": "Combined text from all lines",            "top": 194,            "bottom": 246,            "x0": 72.0,            "x1": 540.0045790943997        }    """    # 检查每一行是否包含 'bottom' 和其他关键字段    for line in paragraph_lines:        if 'bottom' not in line:            line['bottom'] = line.get('top', 0)  # 如果没有bottom字段，则使用top字段值进行补充

    text_list = [line['text'] for line in paragraph_lines]    text = " ".join(text_list)    top = paragraph_lines[0]['top']    bottom = paragraph_lines[-1]['bottom']  # 获取最后一行的bottom值    x0_list = [line['x0'] for line in paragraph_lines]    x1_list = [line['x1'] for line in paragraph_lines]    x0 = min(x0_list)    x1 = max(x1_list)

    return {        "text": text,        "top": top,        "bottom": bottom,        "x0"




    
: x0,        "x1": x1    }

def extract_text_by_paragraph(pdf_path, page_width_threshold=0.85, line_width_difference_ratio=0.05, tolerance=3):    """    基于行的分段和合并策略，按段落提取 PDF 内容。

    Args:        pdf_path (str): PDF 文件路径。        page_width_threshold (float): 页面宽度占比阈值（0~1），超过该比例的行视为段落。

    Returns:        list: 包含每页段落的列表，每个段落包含文字及排版信息。        :param line_width_difference_ratio:    """

    extracted_data = []

    with pdfplumber.open(pdf_path) as pdf:        for page_num, page in enumerate(pdf.pages):            page_width = page.width  # 当前页的总宽度            lines = []

            # 第一步：按行提取文字            for word_info in page.extract_words():                # 将同一行的文字合并                top = round(word_info["top"])  # 取整数以避免浮点误差                bottom = round(word_info["bottom"])                if not lines or abs(lines[-1]["top"] - top) > tolerance:                    # 如果是新的一行，创建新的行数据                    lines.append({                        "text": word_info["text"],                        "top": top,                        "bottom": bottom,                        "x0": word_info["x0"],                        "x1": word_info["x1"],                    })                else:                    # 同一行的文字合并                    lines[-1]["text"] += f" {word_info['text']}"                    lines[-1]["x0"] = min(lines[-1]["x0"], word_info["x0"])  # 更新行的左边界                    lines[-1]["x1"] = max(lines[-1]["x1"], word_info["x1"])  # 更新行的右边界

            # print(json.dumps(lines, indent=4, ensure_ascii=False))

            for word_info in page.extract_words():                # 跳过表格中的文字                if page.find_tables():                    for table in page.find_tables():                        if table.bbox[1] <= word_info["top"] <= table.bbox[3]:                            break  # 属于表格内容，跳过                    else:                        # 非表格内容，合并行文字                        top = round(word_info["top"])                        if not lines or lines[-1]["top"] != top:                            lines.append({                                "text": word_info["text"],                                "top": top,                                "x0": word_info["x0"],                                "x1": word_info["x1"]                            })                        else:                            lines[-1]["text"] += f" {word_info['text']}"

            # 第二步：按段落合并行            paragraphs = []            paragraph_lines = []            for i, line in enumerate(lines):                is_last_line = i == len(lines) - 1  # 判断是否是最后一行                line_width_ratio = line["x1"] / page_width  # 当前行宽度占比

                if i == 0:  # 第一行                    # 段落的起始                    paragraph_lines.append(line)                    if (                            not is_last_line                            and abs(lines[i + 1]["x1"] - line["x1"]) / line["x1"] > line_width_difference_ratio                    ):  # 段落结束                        paragraph = {"page_num": page_num + 1}                        paragraph.update(get_paragraph_lines_info(paragraph_lines))                        paragraphs.append(paragraph)                        paragraph_lines = []                        continue                elif (  # 中间行                        i > 0 and                        not is_last_line                ):                    if (line_width_ratio > page_width_threshold                            and abs(lines[i - 1]["x1"] - line["x1"]) / line["x1"] > line_width_difference_ratio):                        # 段落的起始                        paragraph_lines = [line]                        if (                                not is_last_line                                and abs(lines[i + 1]["x1"] - line["x1"]) / line["x1"] > line_width_difference_ratio                        ):  # 段落结束                            paragraph = {"page_num": page_num + 1}                            paragraph.update(get_paragraph_lines_info(paragraph_lines))                            paragraphs.append(paragraph)                            paragraph_lines = []                            continue                    elif (                            not is_last_line                            and line_width_ratio > page_width_threshold                            and abs(lines[i - 1]["x1"] - line["x1"]) / line["x1"] < line_width_difference_ratio                    ):                        # 当前行是段落的中间部分，合并到段落中                        paragraph_lines.append(line)                    else:                        # 当前行结束段落                        paragraph_lines.append(line)                        paragraph = {"page_num": page_num + 1}                        paragraph.update(get_paragraph_lines_info(paragraph_lines))                        paragraphs.append(paragraph)                        paragraph_lines = []                elif is_last_line:  # 末尾行                    paragraph_lines = [line]                    paragraph = {"page_num": page_num + 1}                    paragraph.update(get_paragraph_lines_info(paragraph_lines))                    paragraphs.append(paragraph)

            extracted_data.append(paragraphs)

    return




    
 extracted_data

25.2.2 抽取图片

[    {        "page_num": 4,  # 图片在PDF中的第几页        "image_index": 1, # 这是这一页中的第几个图片        "path": "./pdf_translate/output/images/page_4_img_1.png", #图片保存位置        "image_ext": "png",  # 图片格式后缀        "width": 3747,  # 图片宽度        "height": 1520,  # 图片高度        "colorspace": 3,  #颜色空间        "bbox": [   # 图片位置信息            93.54499816894531,            471.3414306640625,            501.7056884765625,            636.9150390625        ]    },    {        "page_num": 12,        "image_index": 1,        "path": "./pdf_translate/output/images/page_12_img_1.png",        "image_ext": "png",        "width": 1782,        "height": 1016,        "colorspace": 3,        "bbox": [            114.41100311279297,            241.2252197265625,            477.2440490722656,            448.093017578125        ]    }]

import fitz  # PyMuPDFimport os
"""针对抽取的图片中的字段进行说明。针对下面这个案例说明：{    "page_num": 4,    "image_index": 1,    "path": "./pdf_translate/output/images/page_4_img_1.png",    "image_ext": "png",    "width": 3747,    "height": 1520,    "colorspace": 3,    "bbox": [        93.54499816894531,        471.3414306640625,        501.7056884765625,        636.9150390625    ]}
1.  page_num:    •  含义: 图片所在的 PDF 页码（从 1 开始计数）。    •  单位: 无单位，整数值。    •  例子: "page_num": 4 表示该图片位于 PDF 文件的第 4 页。    2. image_index:    •  含义: 该图片在页面中的索引。每页上的图片都有一个索引值，用于标识该页中的图片顺序。    •  单位: 无单位，整数值。    •  例子: "image_index": 1 表示这是该页面上的第一张图片。    3. path:    •  含义: 提取的图片文件保存路径。该路径指向提取并保存的图片文件。    •  单位: 无单位，字符串表示文件路径。    •  例子: "path": "./pdf_translate/output/images/page_4_img_1.png" 表示该图片保存在 ./pdf_translate/output/images/ 目录下，文件名为 page_4_img_1.png。    4. image_ext:    •  含义: 提取图片的文件扩展名（即图片的格式）。如 png、jpg、jpeg 等。    •  单位: 无单位，字符串表示文件扩展名。    •  例子: "image_ext": "png" 表示该图片被保存为 PNG 格式。    5. width:    •  含义: 图片的宽度，单位是 像素 (px)。这是提取图片的实际分辨率，表示图片的水平像素数量。    •  单位: 像素（px）。    •  例子: "width": 3747 表示图片的宽度为 3747 像素。    6. height:    •  含义: 图片的高度，单位是 像素 (px)。这是提取图片的实际分辨率，表示图片的垂直像素数量。    •  单位: 像素（px）。    •  例子: "height": 1520 表示图片的高度为 1520 像素。    7. colorspace:    •  含义: 图片的颜色空间。不同的颜色空间编码方式定义了图片颜色的表现方式。    •  1 表示 灰度（Grayscale），通常是单色图像。    •  3 表示 RGB 颜色空间（Red, Green, Blue），常见于彩色图片。    •  其他值也代表不同的颜色空间（如 CMYK）。    •  单位: 无单位，整数值，代表颜色空间的编码方式。    •  例子: "colorspace": 3 表示该图片使用 RGB 颜色空间。    8. bbox:    •  含义: 图片在 PDF 页面中的边界框（bounding box），定义了图片在页面中的位置和尺寸。    •  [x0, y0, x1, y1]：x0, y0 表示图片的左下角坐标，x1, y1 表示图片的右上角坐标。    •  这些值是基于 PDF 页面坐标系，单位是 点 (pt)，PDF 的标准单位。1 点 = 1/72 英寸。    •  这个边界框定义了图片在页面中的位置和图片的尺寸。图片的显示区域由此边界框决定。    •  单位: 点（pt），PDF 的标准坐标单位。    •  例子:        "bbox": [        93.54499816894531,        471.3414306640625,        501.7056884765625,        636.9150390625    ]        •   93.54499816894531 是图片左下角的 x 坐标。    •  471.3414306640625 是图片左下角的 y 坐标。    •  501.7056884765625 是图片右上角的 x 坐标。    •  636.9150390625 是图片右上角的 y 坐标。            单位说明

    •  像素 (px):    •  用于 width 和 height 字段，表示图片的分辨率。像素是图像显示的基本单位，定义了图像的精度和清晰度。    •  点 (pt):    •  用于 bbox 字段，表示图片在 PDF 页面中的坐标位置。PDF 坐标系统以点为单位，1 点等于 1/72 英寸。bbox 提供了图片在 PDF 页面中的具体位置和大小。    """

def extract_images(pdf_path, output_dir):    """    从 PDF 提取嵌入图片，并返回正确的边界框 (bbox)。

    Args:        pdf_path (str): PDF 文件路径。        output_dir (str): 保存图片的目录。

    Returns:        list: 包含图片路径、分辨率、颜色空间和位置的元数据列表。    """    os.makedirs(output_dir, exist_ok=True)    image_metadata = []

    pdf_document = fitz.open(pdf_path)

    for page_num in range(len(pdf_document)):        page = pdf_document[page_num]        images = page.get_images(full=True)

        # 遍历页面中的每一张图片        for img_index, img in enumerate(images):            xref = img[0]  # 图片引用号            base_image = pdf_document.extract_image(xref)            image_bytes = base_image["image"]            image_ext = base_image["ext"]

            # 提取图片的 bbox            bbox = None            for img_info in page.get_images(full=True):                if img_info[0] == xref:                    bbox = page.get_image_bbox(img_info)

            # 保存图片            image_filename = f"page_{page_num + 1}_img_{img_index + 1}.{image_ext}"            image_path = os.path.join(output_dir, image_filename)            with open(image_path, "wb") as image_file:                image_file.write(image_bytes)

            # 记录图像元数据            image_metadata.append({                "page_num": page_num + 1,                "image_index": img_index + 1,                "path": image_path,                "image_ext": image_ext,                "width": base_image["width"],                "height": base_image["height"],                "colorspace": base_image["colorspace"],                "bbox": list(bbox) if bbox else None,  # 确保 bbox 是列表格式            })
    pdf_document.close()    return image_metadata

25.2.3 文本翻译

import sysimport jsonfrom tqdm import tqdmsys.path.append('./')from openai import OpenAIfrom pdf_translate.configs.model_config import DEEPSEEK_API_KEY, DEEPSEEK_MODEL, DEEPSEEK_BASE_URL
client = OpenAI(    api_key=DEEPSEEK_API_KEY,    base_url=DEEPSEEK_BASE_URL,)

# 子函数1：生成标题def generate_en_2_zh_translate(content):    """    使用 GPT 生成吸引眼球的中文标题。
    :param content: 英文的文章内容    :return: 生成的中文标题    """    prompt = (        "你是一名专业的翻译，擅长将英文文档翻译成中文。"        "以下是给你翻译的英文原文：\n"        f"{content}\n"        "\n请将上面英文翻译成语句通顺、结构合理、符合中国人表达习惯的中文。注意，如果是阿拉伯数字请不用翻译为中文，英文名、链接等也不用翻译。"    )    response = client.chat.completions.create(        model=DEEPSEEK_MODEL,        messages=[            {"role": "system", "content": "你是一名专业的英文翻译，你擅长将英文翻译为中文。"},            {"role": "user", "content": prompt},        ]    )    message_content = response.choices[0].message.content.strip()    return message_content

def translate_all_pages(text_with_layout, output_path):    """    主函数：翻译整个文档。    """    # 1. 翻译每一页的文字    translated_data = []    for page_data in tqdm(text_with_layout, total=len(text_with_layout), desc=f"translate pages", unit="pages"):        page_translated = []        for paragraph in tqdm(page_data, total=len(page_data), desc=f"translate paragraphs", unit="paragraph"):            original_text = paragraph["text"]            translated_text = generate_en_2_zh_translate(original_text)            paragraph["text"] = translated_text            page_translated.append(paragraph)        translated_data.append(page_translated)
    # 打开文件（如果文件不存在，将自动创建）    with open(output_path, 'w', encoding='utf-8') as file:        file.write(json.dumps(translated_data, indent=4, ensure_ascii=False))

25.2.4 绘制带文本的PDF