xiaomoguhzz/bagel_sharegpt4o

🧠 BLIP3O & ShareGPT4O 数据集（Bagel-Think 子集）

---

📖 数据集简介

数据集	数据来源	任务类型	样本数量	每样本推理结果数	图像总量
BLIP3O Subset	从原始 BLIP3O 数据集中随机采样	文本到图像生成（Text-to-Image）	20,000	各 3（步长 = 3, 5, 7）	60,000
ShareGPT4O Subset	从 ShareGPT4O 图像编辑数据集中采样	图像编辑（Image Editing）	20,000	各 3（步长 = 3, 5, 7；含源图各自复制）	120,000

两者均由 Bagel 框架 推理生成，并采用 泰勒加速 (Taylor Acceleration) 技术，用于研究不同步长下生成速度与质量的平衡关系。

---

🧩 数据生成流程

① 抽样阶段

• BLIP3O：从原始 BLIP3O 数据集中随机抽取 20,000 条文本指令（Prompt）。
• ShareGPT4O：从 ShareGPT4O 图像编辑数据集中选取 20,000 条编辑样本（含原图与编辑指令）。

---

② 推理阶段

两者均在 Bagel 框架下使用三组泰勒加速步长：
[3, 5, 7]

• 步长越少 → 推理速度越快
• 步长越多 → 图像质量通常更高

BLIP3O：直接基于文本生成图像。
ShareGPT4O：针对每个样本执行编辑推理（含同一原图与三组编辑结果）。

---

③ 数据量与目录结构

数据集	图像数量	文件结构说明
BLIP3O	60,000 张	blip3o_bagel_dataset.jsonl images/
ShareGPT4O	120,000 张	sharegpt4o_bagel_dataset.jsonl images/ sharegpt4o_00.tar.gz ~ sharegpt4o_18.tar.gz

🧮 自动评分体系

两个数据集的图像均由 Qwen-VL 2.5-72B 模型自动评估。 但由于任务不同，采用了两套相应的 VLM 评分模板。

🌟 BLIP3O — 文本到图像评估模板

### RULES of this evaluation task:
One image will be provided; the image is AI-generated.
A corresponding text instruction (prompt) will also be given.

The objective is to evaluate:
1. Semantic Consistency — how well the image follows and matches the text instruction.
2. Perceptual Quality — how natural, coherent, and artifact-free the image looks.

### SCORING
Each score ranges from 0 to 10:

(1) Semantic Consistency:
0 → The image is completely unrelated or contradicts the text description.
10 → Every object, relation, and context in the prompt is accurately represented.

(2) Perceptual Quality:
0 → The image appears heavily distorted or unrealistic, with strong artifacts.
10 → The image looks realistic, natural, and visually coherent.

### OUTPUT FORMAT
Output your evaluation strictly as valid JSON:
{
  "semantic_consistency": <0-10>,
  "perceptual_quality": <0-10>,
  "reasoning": "A short summary (30-60 tokens) mentioning main alignment or quality issues."
}

### TEXT INSTRUCTION TO EVALUATE
<T2I_PROMPT>

Please think carefully before judging.
Output only the final JSON object, without any commentary or additional text.

🌟 ShareGPT4O — 图像编辑评估模板

### RULES of this evaluation task:
You will be provided with two images and a corresponding text instruction.
- The first image is the **original image (before editing)**.
- The second image is the **edited image (after applying the editing instruction)**.
- The text instruction describes the intended edit or modification.

The objective is to evaluate:
1. **Edit Consistency** — how accurately the edited image reflects the intended changes described in the instruction, while preserving unrelated content from the original image.
2. **Perceptual Quality** — how natural, coherent, and artifact-free the edited image looks.

### SCORING
Each score ranges from 0 to 10:

(1) Edit Consistency:
0 → The edit is completely unrelated or contradicts the text instruction.
10 → The edit precisely follows the described modification; unchanged areas remain consistent with the original image.

(2) Perceptual Quality:
0 → The edited image appears heavily distorted or unrealistic, with strong artifacts.
10 → The edited image looks realistic, natural, and visually coherent.

### OUTPUT FORMAT
Output your evaluation strictly as valid JSON:
{
  "edit_consistency": <0-10>,
  "perceptual_quality": <0-10>,
  "reasoning": "A short summary (30-60 tokens) mentioning main edit inconsistencies or visual problems only."
}

### EDIT INSTRUCTION TO EVALUATE
<EDIT_PROMPT>

Please think carefully before judging.
Output only the final JSON object, without any commentary or additional text.

📁 数据格式说明

数据集	主文件	字段说明
BLIP3O	blip3o_bagel_dataset.jsonl	image: 图像文件名，例如 "blip3o_gen_fresh3_20000_00000.png" prompt: 输入的文本指令 semantic_consistency: 语义一致性得分 (0–10) perceptual_quality: 感知质量得分 (0–10) reasoning: 模型给出的评分理由 id: 样本编号
ShareGPT4O	sharegpt4o_bagel_dataset.jsonl	src_img: 原始图像文件名 tgt_img: 编辑后图像文件名 prompt: 编辑任务指令 edit_consistency: 编辑一致性得分 (0–10) perceptual_quality: 编辑后图像质量得分 (0–10) reasoning: 模型输出的简要评价轨迹 id: 样本编号

---

📄 示例样本

🔹 BLIP3O 示例

{
  "image": "blip3o_gen_fresh3_20000_00000.png",
  "prompt": "Hacksaw",
  "semantic_consistency": 5,
  "perceptual_quality": 8,
  "reasoning": "The image shows a character with a large saw-like weapon, partially matching 'hacksaw,' but lacks direct textual alignment. Visual quality is high, with few artifacts.",
  "id": 0
}

{
  "src_img": "sharegpt4o_edit_fresh7_20000_00000_src.png",
  "tgt_img": "sharegpt4o_edit_fresh7_20000_00000_tgt.png",
  "prompt": "Convert the teddy bear into a female teddy bear with feminine accessories.",
  "edit_consistency": 9,
  "perceptual_quality": 9,
  "reasoning": "The teddy bear was successfully converted to appear feminine with a pink bow and ribbon, maintaining background consistency and minimal distortion.",
  "id": 0
}