Reasoning対応のOCR特化VLMモデル「NuMarkdown-8B-Thinking」を試す

!pip install "flash-attn==2.7.4.post1" --no-build-isolation

!apt update && apt install -y poppler-utils

!pip install pdf2image

!wget https://www.city.kobe.lg.jp/documents/15123/r5_doukou.pdf

import os
from pdf2image import convert_from_path

kobe_pdf = "r5_doukou.pdf"
kobe_output_dir = "kobe"

def convert_pdf_to_image(pdf_path, output_dir_path):
    os.makedirs(output_dir_path, exist_ok=True)

    images = convert_from_path(pdf_path)
    for i, image in enumerate(images):
        output_path = f"{output_dir_path}/page_{i + 1}.png"
        image.save(output_path, "PNG")
        print(f"Saved: {output_path}")

convert_pdf_to_image(kobe_pdf, kobe_output_dir)

kobe
├── page_10.png
├── page_11.png
├── page_12.png
├── page_13.png
├── page_14.png
├── page_15.png
├── page_16.png
├── page_17.png
├── page_18.png
├── page_19.png
├── page_1.png
├── page_20.png
├── page_21.png
├── page_2.png
├── page_3.png
├── page_4.png
├── page_5.png
├── page_6.png
├── page_7.png
├── page_8.png
└── page_9.png

import torch
from PIL import Image
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "numind/NuMarkdown-8B-reasoning"       

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
    min_pixels=100*28*28, max_pixels=5000*28*28   
)

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
    trust_remote_code=True,
)

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.54.15              Driver Version: 550.54.15      CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA A100-SXM4-40GB          Off |   00000000:00:04.0 Off |                    0 |
| N/A   36C    P0             52W /  400W |   16241MiB /  40960MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

img = Image.open("kobe/page_1.png").convert("RGB")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
        ],
    }
]
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_input = processor(text=prompt, images=[img], return_tensors="pt").to(model.device)

with torch.no_grad():
    model_output = model.generate(**model_input, temperature = 0.7, max_new_tokens=5000)

result = processor.decode(model_output[0])
reasoning = result.split("<think>")[1].split("</think>")[0]
answer  = result.split("<answer>")[1].split("</answer>")[0]
print(answer)

# 令和５年度 神戸市観光動向調査結果について

**調査目的**
本調査は，来神観光客の特質と神戸市内の観光動向を把握し，今後の観光行政の参考とすることを目的に実施。

**調査方法**
調査員が対面にてアンケート項目を聞き取りする形で実施。

**調査日**
① 令和５年 5月 13 日(土)・20 日(土)※
② 令和５年 9月 9 日(土)
③ 令和５年 11月 25 日(土)/12 月 2 日(土)※
④ 令和６年 2月 10 日(土)・17 日/3月 30 日(土)※

※雨天等により、一部の調査地点については延期して実施。

**調査地点**

| 地区 | NO | 調査地点 | サンプル数 |
| :--- | :- | :------- | ----------: |
| 北野 | 1  | 北野町広場 | 246         |
|      | 2  | 北野工房のまち | 154         |
| 市街地 | 3  | 南京町広場 | 218         |
|      | 4  | 総合インフォメーションセンター | 198         |
|      | 5  | 旧居留地 | 201         |
|      | 6  | 神戸布引ロープウェイハーブ園山麓駅 | 203         |
|      | 7  | 白鶴酒造資料館内 | 205         |
|      | 8  | 王子動物園内 | 235         |
|      | 9  | 県立美術館前 | 217         |
|      | 10 | さんちか夢広場 | 208         |
|      | 11 | 神戸空港屋上階展望ロビー | 212         |
|      | 12 | IKEA神戸 | 192         |
|      | 13 | 能福寺（兵庫大仏） | 147         |
|      | 14 | 兵庫津ミュージアム | 204         |
| 神戸港 | 15 | 神戸海洋博物館前及び周辺 | 200         |
|      | 16 | 中突堤中央ターミナル周辺 | 208         |
|      | 17 | umie モザイク | 201         |
|      | 18 | ハーバーランド総合インフォメーション前 | 213         |
| 六甲・摩耶 | 19 | 六甲天覧台 | 209         |
|      | 20 | 摩耶山掬星台広場 | 200         |
|      | 21 | 六甲山牧場内 | 215         |
|      | 22 | 六甲ガーデンテラス内 | 211         |
| 有馬 | 23 | 有馬温泉観光総合案内所前 | 214         |
|      | 24 | 金の湯前 | 216         |
|      | 25 | 六甲有馬ロープウェイ有馬温泉駅 | 232         |
| 須磨・舞子 | 26 | 須磨離宮公園 | 202         |
|      | 27 | 舞子公園（舞子海上プロムナード） | 230         |
| 西北神 | 28 | 神戸ワイナリー（農業公園） | 235         |
|      | 29 | 神戸フルーツ・フラワーパーク大沢園内 | 209         |
|      | 30 | 神戸三田プレミアム・アウトレット | 215         |
| 合計 |   |          | 6,250       |

print(reasoning)

1. **Identify and Transcribe Header**: The document has a main title at the top: "令和５年度 神戸市観光動向調査結果について". This should be rendered as an H1 heading.

2. **Process Introductory Paragraphs**: Below the main title, there are three distinct sections with blue rounded labels: "調査目的", "調査方法", and "調査日". Each label is followed by a paragraph of text. These can be represented as a list or just separate paragraphs. Given their visual layout, treating them as separate paragraphs under their respective labels (which can be bolded or simply part of the text) is appropriate.

3. **Handle "調査日" Section**: This section contains a numbered list of dates. The accompanying footnote "※雨天等により、一部の調査地点については延期して実施。" should be placed directly after the list.

4. **Identify and Convert Main Table Structure**: The majority of the content is a large table titled "調査地点". This table has four columns: "地区", "NO", "調査地点", and "サンプル数".

5. **Determine Table Columns and Headers**: The table clearly has 4 columns. The headers are "地区", "NO", "調査地点", and "サンプル数".

6. **Handle Merged Cells in "地区" Column**: The "地区" column has merged cells for categories like "北野", "市街地", "神戸港", "六甲・摩耶", "有馬", "須磨・舞子", and "西北神". GitHub Flavored Markdown (GFM) tables do not directly support merged cells. The best approach is to repeat the "地区" name for each row it applies to, or leave it blank for subsequent rows within the same "region" if it makes more sense contextually. In this case, repeating the region name is clearer and maintains the logical grouping.

7. **Transcribe Table Content Row by Row**: Carefully transcribe all text and numbers into the table cells. Pay attention to Japanese characters and special characters like "・" and "/".

8. **Identify and Transcribe Table Footer**: There is a "合計" row at the bottom right, indicating a total sample count. This should be included as a final row in the table.

9. **Final Review**: Check for any missed text, incorrect formatting, or issues with table alignment. Ensure all Japanese text is accurately transcribed. The overall structure is straightforward, so no complex LaTeX or image placeholders are needed beyond the table itself.

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.54.15              Driver Version: 550.54.15      CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA A100-SXM4-40GB          Off |   00000000:00:04.0 Off |                    0 |
| N/A   70C    P0            392W /  400W |   37315MiB /  40960MiB |    100%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

img = Image.open("kobe/page_4.png").convert("RGB")
(snip)

# 分析の詳細

## (1) 性別構成

- 性別構成としては、全市では「男性」が43.5%、「女性」が56.5%となっている。
- 地区別にみると、『六甲・摩耶』を除いた地区で女性の占める割合が高いが、『北野』では「女性」が30.0ポイント高く、差が大きくなっている。

【図表 1　性別構成】

| (%) | 全市 | 北野 | 市街地 | 神戸港 | 六甲・摩耶 | 有馬 | 須磨・舞子 | 西北神 |
|---|---|---|---|---|---|---|---|---|
| 男性 | 43.5 | 35.0 | 42.4 | 41.4 | 53.8 | 42.4 | 42.8 | 44.3 |
| 女性 | 56.5 | 65.0 | 57.6 | 58.6 | 46.2 | 57.6 | 57.2 | 55.7 |

![Bar chart showing male and female percentages for Fiscal Year 5 and Fiscal Year 4, with categories for Kita-nokyo, Shibuichi, Kobe Port, Rokko-Moye, Yumoto, Suma-Wakasagi, and Nishikyogoku regions.](image_1.png)

## (2) 年齢構成

- 年齢構成としては、全市では「60歳以上」が24.3%で最も高く、次いで「50歳代」(19.8%)、「40歳代」(18.0%)、「20歳代」(16.9%)、「30歳代」(16.2%)の順となっており、40歳以上が6割を超える。
- 地区別にみると、『須磨・舞子』では「60歳以上」(35.9%)が他の地区と比較して高くなっている。

【図表 2　年齢構成】

| (%) | 全市 | 北野 | 市街地 | 神戸港 | 六甲・摩耶 | 有馬 | 須磨・舞子 | 西北神 |
|---|---|---|---|---|---|---|---|---|
| 19歳以下 | 4.7 | 6.3 | 3.6 | 5.6 | 5.0 | 6.2 | 2.3 | 6.2 |
| 20歳代 | 16.9 | 24.5 | 13.7 | 18.1 | 23.0 | 22.4 | 12.0 | 12.6 |
| 30歳代 | 16.2 | 13.3 | 17.9 | 17.5 | 15.3 | 13.6 | 15.3 | 14.7 |
| 40歳代 | 18.0 | 14.3 | 18.6 | 18.4 | 18.0 | 14.5 | 16.9 | 22.0 |
| 50歳代 | 19.8 | 20.5 | 19.8 | 18.9 | 20.0 | 20.7 | 17.6 | 21.2 |
| 60歳以上 | 24.3 | 21.3 | 26.4 | 21.5 | 18.7 | 22.7 | 35.9 | 23.2 |

![Bar chart showing age distribution percentages for Fiscal Year 5 and Fiscal Year 4, with age groups from 19 below to 60 and above, and categories for Kita-nokyo, Shibuichi, Kobe Port, Rokko-Moye, Yumoto, Suma-Wakasagi, and Nishikyogoku regions.](image_2.png)

3

1. **Identify and Transcribe Main Sections and Headings:**
   The document is structured with clear headings and subheadings. I will transcribe these directly, using appropriate Markdown heading levels. The main sections are "分析の詳細" (Analysis Details) and its sub-sections "(1) 性別構成" (Sexual Composition) and "(2) 年齢構成" (Age Composition).

2. **Transcribe Introductory Text and Bullet Points:**
   Under each main section, there is introductory text followed by bullet points. I will transcribe this text as plain paragraphs and then convert the bullet points into Markdown list items. The bullet points are indicated by a blue circular icon.

3. **Handle Tables:**
   There are two tables in the document.
   - **Table 1 (Sexual Composition):**
     - The table has a main header "性別構成" (Sexual Composition) which spans multiple columns. This needs to be represented correctly in Markdown.
     - The first column is "性別" (Gender).
     - The second column is "全市" (Overall).
     - The subsequent columns are region names: "北野", "市街地", "神戸港", "六甲・摩耶", "有馬", "須磨・舞子", "西北神".
     - The data is numerical percentages.
     - I will create a Markdown table with these columns.
   - **Table 2 (Age Composition):**
     - Similar structure to Table 1, with "年齢構成" (Age Composition) as the main header.
     - The first column is "年齢" (Age).
     - The second column is "全市" (Overall).
     - The subsequent columns are region names: "北野", "市街地", "神戸港", "六甲・摩耶", "有馬", "須磨・舞子", "西北神".
     - The data is numerical percentages.
     - I will create a Markdown table with these columns.

4. **Handle Bar Charts:**
   There are two bar charts in the document.
   - **Bar Chart 1 (Sexual Composition):**
     - This chart shows "男性" (Male) and "女性" (Female) percentages for "令和5年度 (全市)" (Fiscal Year 5 (Overall)) and "令和4年度 (全市)" (Fiscal Year 4 (Overall)).
     - The x-axis is labeled "％" (Percent).
     - The y-axis has labels for the years and the overall category.
     - The values are displayed inside the bars.
     - Since Markdown does not support embedded images or complex diagrams like bar charts, I will represent them as image placeholders with a descriptive alt text.
   - **Bar Chart 2 (Age Composition):**
     - This chart shows age group percentages for "令和5年度 (全市)" (Fiscal Year 5 (Overall)) and "令和4年度 (全市)" (Fiscal Year 4 (Overall)).
     - The x-axis is labeled "％" (Percent).
     - The y-axis has labels for the years and the overall category.
     - The age groups are "19歳以下", "20歳代", "30歳代", "40歳代", "50歳代", "60歳以上".
     - The values are displayed inside the bars.
     - Again, I will represent this as an image placeholder.

5. **Transcribe Page Number:**
   The document has a page number "3" at the bottom center. I will include this at the very end of the Markdown output.

6. **Review and Format:**
   - Ensure all text is transcribed accurately.
   - Check Markdown syntax for tables (pipes, hyphens for headers, correct alignment).
   - Verify list item formatting.
   - Ensure image placeholders are correctly named and ordered.
   - Maintain the original document's logical flow.

Thu Aug  7 18:38:56 2025       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.54.15              Driver Version: 550.54.15      CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA A100-SXM4-40GB          Off |   00000000:00:04.0 Off |                    0 |
| N/A   36C    P0             52W /  400W |   39515MiB /  40960MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+