Ollama视觉模型实测

我简单地拍了几张营养标签的照片，并使用提示转换为 JSON。然后我针对一堆自托管模型和 Mistral Open API 进行了测试。它们都表现得相当好，但 Mistral API 是最好的。

你知道，击败基准测试是一回事。但基准测试通常是自我报告的随机变量，就像驾照考试和驾驶并不完全一样，它们不能保证给定的现实生活问题会以可行的方式解决。这就是为什么我想给它一个真实的测试用例。

测试很简单：我拍了一堆营养标签的照片，就像下面这张：

视觉模型在将其转换为 JPEG 方面有多好？

我尝试提取以下格式的 JSON：

复制代码

{
  "calories": 180,
  "serving_size": 28.0,
  "unit": "g"
}

我使用的提示：

复制代码

The image is the label of a packaged food product.

Extract the following information from the nutrition facts label and return ONLY a JSON object with no additional text:

- calories: total calories per serving (integer)
- serving_size: numeric serving size value (float)
- unit: unit of measurement for serving size (string, e.g., "g", "ml", "oz", "cup")

Use this exact format:
```json
{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "type": "object",
  "properties": {
    "calories": {
      "type": "integer",
      "description": "Total calories per serving"
    },
    "serving_size": {
      "type": "number",
      "description": "Numeric serving size"
    },
    "unit": {
      "type": "string",
      "description": "Unit of measurement (e.g., g, ml, oz, cups)"
    }
  },
  "required": ["calories", "serving_size", "unit"]
}

Example:

json 复制代码

{
  "calories": 180,
  "serving_size": 28.0,
  "unit": "g"
}

If any value cannot be determined from the label, use null for that field.

复制代码

我运行同样的东西 50 次，看看是否能得到期望的结果。

我使用的评估标准：

prompt_path: ".../prompts/calories_and_serving_size.md"

repeat: 50

threshold: 0

cases:

id: calories-label-180

steps:
- input:
  image_path: "images/IMG_B768CE83-9FEC-461A-BE63-CDDF64EBEB58.jpeg"
  max_tokens: 64
  expectations:
  - type: equals
    value:
    calories: 180
    serving_size: 40.0
    unit: g
id: calories-label-harvest-trail-mix-1

steps:
- input:
  image_path: "images/IMG_3236.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 220
      serving_size: 50.0
      unit: g
    - calories: 220
      serving_size: 0.5
      unit: cup
    - calories: 220
      serving_size: 50.0
      unit: grams
id: calories-label-chunky-supreme-granola-exact

steps:
- input:
  image_path: "images/IMG_3228.HEIC"
  max_tokens: 64
  expectations:
  - type: equals
    value:
    calories: 630
    serving_size: 140.0
    unit: g
id: calories-label-maple-eh-granola-exact

steps:
- input:
  image_path: "images/IMG_3231.HEIC"
  max_tokens: 64
  expectations:
  - type: equals
    value:
    calories: 570
    serving_size: 130.0
    unit: g
id: calories-label-harvest-trail-mix-2

steps:
- input:
  image_path: "images/IMG_3232.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 220
      serving_size: 50.0
      unit: g
    - calories: 220
      serving_size: 0.5
      unit: cup
    - calories: 220
      serving_size: 50.0
      unit: grams
id: calories-label-harvest-trail-mix-3

steps:
- input:
  image_path: "images/IMG_3237.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 220
      serving_size: 50.0
      unit: g
    - calories: 220
      serving_size: 0.5
      unit: cup
    - calories: 220
      serving_size: 50.0
      unit: grams
id: calories-label-hersheys-christmas-cookies-exact

steps:
- input:
  image_path: "images/IMG_3276.HEIC"
  max_tokens: 64
  expectations:
  - type: equals
    value:
    calories: 200
    serving_size: 41.0
    unit: g
id: calories-label-mnm-mint-chocolate-candy-1

steps:
- input:
  image_path: "images/IMG_3279.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 190
      serving_size: 40.0
      unit: g
    - calories: 220
      serving_size: 0.25
      unit: cup
id: calories-label-hersheys-candy-cane-kisses

steps:
- input:
  image_path: "images/IMG_3280.HEIC"
  max_tokens: 64
  expectations:
  - type: equals
    value:
    calories: 220
    serving_size: 42.0
    unit: g
id: calories-label-milk-chocolate-christmas-balls

steps:
- input:
  image_path: "images/IMG_3282.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 200
      serving_size: 40.0
      unit: g
    - calories: 200
      serving_size: 6.0
      unit: pcs
    - calories: 200
      serving_size: 6.0
      unit: pieces
id: calories-label-mnm-mint-chocolate-candy-2

steps:
- input:
  image_path: "images/IMG_3284.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 190
      serving_size: 40.0
      unit: g
    - calories: 220
      serving_size: 0.25
      unit: cup
id: calories-label-jumbo-sour-sukkers

steps:
- input:
  image_path: "images/IMG_3286.HEIC"
  max_tokens: 64
  expectations:
  - type: equals
    value:
    calories: 150
    serving_size: 40.0
    unit: g
  - type: oneOf
    values:
    - calories: 150
      serving_size: 40.0
      unit: g
    - calories: 150
      serving_size: 2.0
      unit: pcs
    - calories: 150
      serving_size: 2.0
      unit: pieces
id: calories-label-quality-street

steps:
- input:
  image_path: "images/IMG_3288.HEIC"
  max_tokens: 64
  expectations:
  - type: oneOf
    values:
    - calories: 200
      serving_size: 42.0
      unit: g
    - calories: 200
      serving_size: 5.0
      unit: pcs
    - calories: 200
      serving_size: 5.0
      unit: pieces
  不再废话，以下是结果：
  
  测试用例的平均模型准确度让我们分解一下：我认为 Llava 失败了。我认为 Llava 是一个 OCR 模型，旨在像 DeepSeek-OCR 一样使用。这些模型不属于这里，它们不遵循像"以 JSON 格式输出"这样的具体指令。
  
  第二个学习是：Qwen3，8B 参数，在 16GB-VRAM 机器上工作得很好。但是，在现实设置中，它们可能并没有比在更小机器上也能工作的 Qwen3:2B 带来太多 ------ 它们可能即使在 6GB VRAM 上也能正常工作。使用 8B 来实现与统计噪音无法区分的性能提升可能是不可行的。
  
  另一个观点是，Qwen3:2B 似乎比 Qwen3:4B 更好。我不知道这是否真的是改进，或者只是统计波动。我目前的想法是：如果应用程序实际上需要一个小窗口，那么增加上下文窗口确实没有意义。不仅会增加成本，实际上可能还会降低性能。
  
  最后一个学习：托管 Ollama 时，打开流式传输！否则，当请求超过 60 秒时，请求会被丢弃。我听说这是可定制的，但默认为流式传输可能更容易，尽管需要额外的代码。
  
  原文链接：Ollama视觉模型实测 - 汇智网