KnowCP Dataset Repository

Project homepage and benchmark details: https://41-edu.github.io/KnowCP-Benchmark/

Script repository: https://github.com/41-edu/KnowCP

Project Overview

KnowCP is a comprehensive benchmark designed to evaluate multimodal large language models on the understanding of Chinese paintings, ranging from basic recognition tasks to deep reasoning challenges.

This repository serves as the dataset repository for KnowCP. It stores the image corpus, annotation resources, question sets, and knowledge-base metadata utilized by the benchmark.

Dataset Scale

The following statistics are aligned with the configuration of the benchmark website and public content files.

1. Paintings: 1210 artworks, some of which consist of multiple pages, resulting in a total of 2331 images.
2. Seal Annotation: 8690 annotations, corresponding to seals on 1792 images, so we constructed 1792 seal recognition questions.
3. Inscription Annotation: 2434 annotations, corresponding to inscriptions on 2351 images, so we constructed 2351 inscription recognition questions.
4. Element Annotation: 5048 annotations.
5. Technique Annotation: 1312 annotations.
6. Deep Reasoning:
   • Visual Analysis: 826 questions
   • Cultural Context: 716 questions
   • Provenance Research: 834 questions
7. Total question items used by the benchmark: 24700

Repository Contents

1. images: All painting images and related sub-images used by the benchmark.

2. questions: All benchmark QA files. We provide 14 question files under questions/by_type, with each file corresponding to one question source type. All QA content is in Chinese. For the specific presentation style of QA and English-facing examples, refer to Cases in https://41-edu.github.io/KnowCP-Benchmark/#distribution

   • Foundational Knowledge

     • Image-to-Title (ITT): Determine the title of a painting composed of a single page.

       • ITT data is stored in questions/by_type/ITT_MHQA_choice.jsonl and questions/by_type/ITT_MHQA_fillin.jsonl

       • Example:

         • plaintext

           {"qid": "IMG000003-P1-Q1", "image_id": "IMG000003", "phase": "P1", "question_no": "Q1", 
            "type": "ITT", "answer_format": "single_choice", 
            "prompt": "请判断该作品标题（中英文对照）是以下哪一项？ 
                A. 白鸽 / Pigeon 
                B. 江村秋晓图 / River Village in Autumn Dawn 
                C. 冬景山水图 / Wintry Mountains 
                D. 画天中华瑞 / Beautiful Scenery of the Tianzhong 
           请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
            "ground_truth": "C", 
            "image_paths": ["images/IMG000003_0.jpg"]}
           

           • "type": "ITT" for ITT questions
           • The image uploaded during runtime is specified in "image_paths": ["images/IMG000003_0.jpg"]
           • The question content is in "prompt"
           • The expected answer is in "ground_truth"

     • Multi-Image-to-Title (MITT): Determine the title of a painting composed of multiple pages.

       • MITT data is stored in questions/by_type/MITT_MHQA_choice.jsonl and questions/by_type/MITT_MHQA_fillin.jsonl

       • Example:

         • plaintext

           {"qid": "IMG000002-P1-Q1", "image_id": "IMG000002", "phase": "P1", "question_no": "Q1", 
            "type": "MITT", "answer_format": "single_choice", 
            "prompt": "请综合观察这些同一作品的多张图像，判断其标题（中英文对照）是以下哪一项？ 
                A. 春郊游骑图 / A Horseback Trip in a Spring Countryside 
                B. 春山青霁 / Clear Skies on Spring Mountain 
                C. 螽斯绵瓞 / Melon and Katydid 
                D. 山水图（十六开） / Landscapes with Poems (Sixteen Leaves) 
           请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
            "ground_truth": "D", 
            "image_paths": ["images/IMG000002_0.jpg", "images/IMG000002_1.jpg", "images/IMG000002_2.jpg"]}
           

           • "type": "MITT" for MITT questions
           • The images uploaded during runtime are specified in "image_paths": ["images/IMG000002_0.jpg", "images/IMG000002_1.jpg", "images/IMG000002_2.jpg"]
           • The question content is in "prompt"
           • The expected answer is in "ground_truth"

     • Title-to-Image (TTI): Select the image that matches a given title from multiple images.

       • Example:

         • plaintext

           {"qid": "IMG000001-P1-Q6", "image_id": "IMG000001", "phase": "P1", "question_no": "Q6", 
           "type": "TTI", "answer_format": "single_choice", 
           "prompt": "已知作品标题为：山水图（十二开） / Landscapes and Trees (Twelve Leaves)。请判断它对应下列四张图中的哪一张？ 
               A. 图像A 
               B. 图像B 
               C. 图像C 
               D. 图像D 
           请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
           "ground_truth": "A", 
           "image_paths": ["images/IMG000001_0.jpg", "images/IMG000002_0.jpg", "images/IMG000003_0.jpg", "images/IMG000004_0.jpg"]}
           

           • "type": "TTI" for TTI questions
           • The images uploaded during runtime are specified in "image_paths": ["images/IMG000001_0.jpg", "images/IMG000002_0.jpg", "images/IMG000003_0.jpg", "images/IMG000004_0.jpg"]}
           • The question content is in "prompt"
           • The expected answer is in "ground_truth"

     • Multi-hop-QA (MHQA): QA on the metadata of each painting.

       • MHQA data is stored in questions/by_type/ITT_MHQA_choice.jsonl, questions/by_type/ITT_MHQA_fillin.jsonl, questions/by_type/MITT_MHQA_choice.jsonl and questions/by_type/MITT_MHQA_fillin.jsonl

       • MHQA includes metadata QA for paintings with multiple pages and single pages. This QA is a sequential question-and-answer process covering four aspects: author, dynasty, museum, and medium. When asking a new question, we append the context of the previous question to the prompt.

       • Example:

         • plaintext

           {"qid": "IMG000003-P1-Q2", "image_id": "IMG000003", "phase": "P1", "question_no": "Q2", 
           "type": "ITT_MHQA_choice", "answer_format": "single_choice", 
           "prompt": "该作品的作者是哪一项？ A. 龚贤 B. 樊圻 C. 吴宏 D. 高岑 请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
           "ground_truth": "A", 
           "image_paths": ["images/IMG000003_0.jpg"]}
           
           {"qid": "IMG000003-P1-Q3", "image_id": "IMG000003", "phase": "P1", "question_no": "Q3", 
           "type": "ITT_MHQA_choice", "answer_format": "single_choice", 
           "prompt": "该作品的朝代是哪一项？ A. 清 B. 晋 C. 金 D. 明 请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
           "ground_truth": "A", 
           "image_paths": ["images/IMG000003_0.jpg"]}
           
           {"qid": "IMG000003-P1-Q4", "image_id": "IMG000003", "phase": "P1", "question_no": "Q4", 
           "type": "ITT_MHQA_choice", "answer_format": "single_choice", 
           "prompt": "该作品的收藏机构为哪一项？ A. 佛利尔·赛克勒美术馆 B. 东京艺术大学 C. 东京国立博物馆 D. 大都会博物馆 请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
           "ground_truth": "D", 
           "image_paths": ["images/IMG000003_0.jpg"]}
           
           {"qid": "IMG000003-P1-Q5", "image_id": "IMG000003", "phase": "P1", "question_no": "Q5", 
           "type": "ITT_MHQA_choice", "answer_format": "single_choice", 
           "prompt": "该作品的材质是哪一项？ A. 棉布 B. 绫本 C. 纸本 D. 金笺 请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
           "ground_truth": "C", 
           "image_paths": ["images/IMG000003_0.jpg"]}
           

         • This is the structure of an entire sequence of MHQA questions. This type includes four sub-types: ITT_MHQA_choice, ITT_MHQA_fillin, MITT_MHQA_choice, and MITT_MHQA_fillin.

         • The following is an example of a prompt with appended context:

           • plaintext

             {"qid": "IMG000003-P1-Q2", "image_id": "IMG000003", "phase": "P1", "question_no": "Q2", 
             "type": "ITT_MHQA_choice", "answer_format": "single_choice", 
             "prompt": "以下是同一作品前序问答，请结合上下文回答当前问题： 
                 第1轮问题： 请判断该作品标题（中英文对照）是以下哪一项？ A. 瓮牖图 / Round Window Scroll B. 秋江钓艇 / Fishing Boat on Autumn River C. 归渔图 / Returning Fishermen D. 冬景山水图 / Wintry Mountains 请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。 
                 第1轮回答：D 
                 当前问题： 该作品的作者是哪一项？ A. 龚贤 B. 樊圻 C. 吴宏 D. 高岑 请只输出选项字母（A/B/C/D）。  输出格式硬约束： - 只输出一个选项字母（如 A）。 - 禁止输出解释、标点、额外文本。", 
             "ground_truth": "A", 
             "image_paths": ["images/IMG000003_0.jpg"], 
             "request_image_paths": ["images\\IMG000003_0.jpg"], "answer": "A", "raw_answer": "A", "answer_structured": {"text": "A"}, "error": "", "skipped": false}
             

   • Visual Content

     • Seal-Recognition (SR)

       • Example

         • plaintext

           {"qid": "IMG000001-P2-Q1-Q01-IMG01", "image_id": "IMG000001", "phase": "P2", "question_no": "Q1", 
           "type": "SR", "answer_format": "json", 
           "prompt": "请检测这张图中的所有印章。 输出必须是 JSON 数组，每个元素代表一个印章。 每个元素至少包含字段：bbox、text。 - bbox: 印章边界框坐标（具体坐标格式将由系统补充说明）。 - text: 该印章印文；无法识别时可填空字符串。 - text 字段请统一输出简体中文；若识别到繁体或异体字，请在不改变语义的前提下尽量转换为简体。 请只输出 JSON，不要输出解释。  输出格式硬约束： - 只输出合法 JSON（对象或数组）。 - 禁止输出 Markdown 代码块、解释性文字。\ \ 坐标协议： bbox 使用 [x_min, y_min, x_max, y_max]，坐标为 0~1 归一化浮点数（voc_xyxy_norm_01）。 输出约束： 1) 仅输出 JSON 数组。 2) 每个元素必须包含 bbox 和 text 字段。 3) 字段名必须是 text，不要使用 text_content / textContent。 4) 坐标必须满足 x_min < x_max 且 y_min < y_max。 5) 未检测到目标时输出 []。", 
           "ground_truth": {"task": "seal_detection", "items": [{"bbox": [0.587156, 0.735297, 0.630989, 0.85107], "text": "印章", "sub_image_path": "images/IMG000001_0.jpg"}, {"bbox": [0.01631, 0.830936, 0.057085, 0.926575], "text": "印章", "sub_image_path": "images/IMG000001_0.jpg"}], "texts_from_seals_json": ["印章"], "image_paths": ["images/IMG000001_0.jpg"]}, 
           "image_paths": ["images/IMG000001_0.jpg"]}
           

         • The question is in "prompt", which requires the model to output the bounding boxes of all seals in the image.

         • The expected answer is in "ground_truth", which contains the annotation results from our annotators as the ground truth.

     • Inscription-Recognition (IR)

       • Example

         • plaintext

           {"qid": "IMG000012-P2-Q2-IMG01", "image_id": "IMG000012", "phase": "P2", "question_no": "Q2", 
           "type": "IR", "answer_format": "json", 
           "prompt": "请检测这张图中的所有题跋/款识文字区域。 输出必须是 JSON 数组，每个元素代表一个题跋区域。 每个元素至少包含字段：bbox、text。 - bbox: 题跋区域边界框（具体坐标格式将由系统补充说明）。 - text: 该区域识别出的题跋文字；无法识别时可填空字符串。 - text 字段请统一输出简体中文；若识别到繁体或异体字，请在不改变语义的前提下尽量转换为简体。 请只输出 JSON，不要输出解释。  输出格式硬约束： - 只输出合法 JSON（对象或数组）。 - 禁止输出 Markdown 代码块、解释性文字。\ \ 坐标协议： bbox 使用 [x_min, y_min, x_max, y_max]，坐标为 0~1 归一化浮点数（voc_xyxy_norm_01）。 输出约束： 1) 仅输出 JSON 数组。 2) 每个元素必须包含 bbox 和 text 字段。 3) 字段名必须是 text，不要使用 text_content / textContent。 4) 坐标必须满足 x_min < x_max 且 y_min < y_max。 5) 未检测到目标时输出 []。", 
           "ground_truth": {"task": "inscription_detection", "items": [{"bbox": [0.076534, 0.037318, 0.199841, 0.249891], "text": "太古蛰龙醒，蚕丛霹雳开。五溪云不去，三峡雪飞来。瞿山清。", "text_gt_available": true, "sub_image_path": "images/IMG000012_0.jpg"}], "kb_inscription": "太古蛰龙醒，蚕丛霹雳开。五溪云不去，三峡雪飞来。瞿山清。", "image_paths": ["images/IMG000012_0.jpg"]}, 
           "image_paths": ["images/IMG000012_0.jpg"]}
           

         • The question is in "prompt", which requires the model to output the bounding boxes of all inscriptions in the image.

         • The expected answer is in "ground_truth", which contains the annotation results from our annotators as the ground truth.

     • Element-Recognition (ER): Provided in fill-in formats.

       • Example

         • plaintext

           {"qid": "IMG000001-P2-Q3_O001_L2_FILLIN", "image_id": "IMG000001", "phase": "P2", "question_no": "Q3_O001_L2_FILLIN",
           "type": "ER_fillin", "answer_format": "text",
           "prompt": "请观察红框中的主体物象，并判断它具体是什么。 请只输出一个最合适的物象名称（例如：松树、拱桥、高士），不要输出路径或解释。  输出格式硬约束： - 仅输出最终答案，不要解释。",
           "ground_truth": "茅屋",
           "image_paths": ["images/IMG000001_4.jpg"], "focus_bbox": [213.36073997944501, 233.20400194911036, 384.7893114080165, 407.0991715277332], "focus_image_index": 5, "focus_sub_image_path": "images/IMG000001_4.jpg"}
           

         • The uploaded materials include the image itself ("image_paths": ["images/IMG000001_4.jpg"]) and a red annotation box ("focus_bbox": [213.36073997944501, 233.20400194911036, 384.7893114080165, 407.0991715277332]).

         • The question is in "prompt", which requires the model to identify the object within the given red box.

         • The expected answer is in "ground_truth".

     • Technique-Recognition (TR): Provided in fill-in formats.

       • Example

         • plaintext

           {"qid": "IMG000001-P3-Q1_R001", "image_id": "IMG000001", "phase": "P3", "question_no": "Q1_R001",
           "type": "TR_fillin", "answer_format": "text",
           "prompt": "请观察红框区域，判断主要绘画技法。 请直接输出技法名称。  输出格式硬约束： - 仅输出最终答案，不要解释。",
           "ground_truth": "披麻皴",
           "image_paths": ["images/IMG000001_0.jpg"], "focus_bbox": [158.2181562604731, 537.6836388751759, 783.4618818175318, 727.151434498527], "focus_image_index": 1, "focus_sub_image_path": "images/IMG000001_0.jpg"}
           

         • The uploaded materials include the image itself ("image_paths": ["images/IMG000001_0.jpg"]]) and a red annotation box ("focus_bbox": [158.2181562604731, 537.6836388751759, 783.4618818175318, 727.151434498527]).

         • The question is in "prompt", which requires the model to identify the painting technique used in the given red box.

         • The expected answer is in "ground_truth".

   • Deep Reasoning

     • Visual-Analysis (VA)

       • Example

         • plaintext

           {"qid": "IMG000002-P4-Q3A", "image_id": "IMG000002", "phase": "P4", "question_no": "Q3A", 
           "type": "VA", "answer_format": "text", 
           "prompt": "请客观描述画面主要内容、主体关系与关键视觉要素。 要求： 1）仅基于图像本身描述，不引入外部史料。 2）控制在 200 字以内。 3）优先覆盖主体、动作、空间关系。  输出格式硬约束： - 仅输出最终答案，不要解释。", 
           "ground_truth": "本册笔墨清雅，描摹精妙，所写奇山异水各具风致。第一开绘一座半岛伸入湖中，三面环水，对岸群山间垂瀑而下，景物安排繁简得当，形成含蓄悠远的空间效果。第三开写近处山泉、石桥，对岸层峦叠嶂，半山有洞门半开，境界如仙。第六开则绘江边高台，四面空阔，山光与清风之感相互生发，景象疏朗高远。  全册山石树木多经反复烘染，呈现出阴阳分明而又朦胧混沌的视觉效果。程正揆所谓“用笔墨如龙驭风，如云行空，隐现变幻”，点出了其笔墨不以雄强取胜、而以韵致见长的特点。这种不可捉摸、恍惚迷幻的笔墨表现，构成了龚贤绘画特有的意境。", 
           "image_paths": ["images/IMG000002_0.jpg", "images/IMG000002_1.jpg", "images/IMG000002_2.jpg"]}
           

         • The question is in "prompt", which requires the model to provide a textual description of the visual information of the painting.

         • The expected answer is in "ground_truth", which is a carefully compiled text by experts in Chinese painting.

     • Cultural-Context (CC)

       • Example

         • plaintext

           {"qid": "IMG000002-P4-Q3B", "image_id": "IMG000002", "phase": "P4", "question_no": "Q3B", 
           "type": "CC", "answer_format": "text", 
           "prompt": "请结合题材来源与图文对应关系，解释这幅作品的文化含义。 要求： 1）说明图像元素与题材文本的对应。 2）控制在 220 字以内。 3）避免泛泛而谈。  输出格式硬约束： - 仅输出最终答案，不要解释。", 
           "ground_truth": "第六开涉及昭明太子题材。昭明太子萧统在文学史上享有盛名，南京、镇江、常熟等地都流传其读书遗迹，虽多有附会，但这一文化记忆为画中“江边高台、四面空阔”的景象提供了可供联想的人文背景，使画面不仅是山水描写，也带有追怀古人、寄托文雅想象的意味。  有研究者对册中诗题作过考释，认为此册“诗画参读，极耐玩味”，体现出画与诗的紧密结合。册后两开画跋又集中阐述龚贤关于“作画”与“识画”关系的理论，强调“未学画先知看画”，将“看画”视为作画精进的重要前提。这些诗题与跋文共同构成了理解此册的重要文本资源，有助于从题咏、画论与观画方法的层面把握作品意涵。", 
           "image_paths": ["images/IMG000002_0.jpg", "images/IMG000002_1.jpg", "images/IMG000002_2.jpg"]}
           

         • The question is in "prompt", which requires the model to provide an explanation of the cultural background of the painting.

         • The expected answer is in "ground_truth", which is a document verified by experts in Chinese painting through multiple sources.

     • Provenance-Research (PR)

       • Example

         • plaintext

           {"qid": "IMG000002-P4-Q3C", "image_id": "IMG000002", "phase": "P4", "question_no": "Q3C", 
           "type": "PR", "answer_format": "text", 
           "prompt": "请概述这幅作品的版本流传、归属争议与断代判断。 要求： 1）覆盖款识/著录/收藏或学界判断中的关键点。 2）控制在 200 字以内。 3）结论需有依据，不要编造来源。  输出格式硬约束： - 仅输出最终答案，不要解释。", "ground_truth": "本册绘于康熙二十七年（1688），为龚贤去世前一年的重要作品之一，在其晚年创作中具有代表性。程正揆称龚贤“盖以韵胜，不以力雄者也”，这一评价也提示了此册在龚贤艺术风格中的位置，即以“韵胜”为区别于同时代画家的重要特征。  册后跋文提及宋代宣和内府、倪瓒清閟阁、顾瑛玉山草堂等历史著名藏画之所，虽主要用于申说“识画”理论，但同时关联到古代鉴藏传统与绘画鉴识史的脉络。据钤印可知，本册曾经金城、张学良等人鉴藏。", 
           "image_paths": ["images/IMG000002_0.jpg", "images/IMG000002_1.jpg", "images/IMG000002_2.jpg"]}
           

         • The question is in "prompt", which requires the model to provide a textual description of the provenance of the painting.

         • The expected answer is in "ground_truth", which is a document verified by experts in Chinese painting through multiple sources.

3. annotations

   Fine-grained annotations produced by our annotators for each painting image, including seals, inscriptions, elements, and techniques.

   • Annotation visualization references: https://41-edu.github.io/KnowCP-Benchmark/#distribution

4. kb

   Core metadata per painting, including identity and background attributes used by benchmark tasks.

If You Want to Run Evaluation

This repository provides data only.

For actual model benchmarking and score computation, please use the script repository:

https://github.com/41-edu/KnowCP