dataset card

README.md

@@ -1,29 +1,143 @@
+---
+pretty_name: KnowCP
+language:
+- zh
+license: cc-by-4.0
+task_categories:
+- visual-question-answering
+- image-text-to-text
+task_ids:
+- open-ended-vqa
+- multi-choice-vqa
+- visual-grounding
+size_categories:
+- 10K<n<100K
+tags:
+- chinese-painting
+- cultural-heritage
+- multimodal
+- benchmark
+---
 # KnowCP
 
-KnowCP is a benchmark dataset for Chinese painting understanding and evaluation.
-
-## Repository links
-- Project website and code: https://github.com/41-edu/KnowCP
-- Dataset page: https://huggingface.co/datasets/g41/KnowCP
-
-## Directory structure
-- metadata/schema.json
-- questions/questions_all.jsonl
-- questions/by_type/*.jsonl
-- images/**
-- prompts/constraint_prompts.json
-
-## Public question fields
-Each public question item should keep only:
-- qid
-- phase
-- question_type
-- question_text
-- ground_truth
-- constraint_prompt_id (optional)
-- question_text_en (optional)
-
-Internal metadata and internal storage paths should be removed in public release.
-
-## License
-CC-BY-4.0
+KnowCP is a benchmark for Chinese painting understanding and reasoning.
+It supports both recognition-style tasks (region/text extraction) and knowledge/reasoning tasks (open QA, multiple choice, and multi-turn QA).
+
+## Dataset Summary
+
+- Name: KnowCP
+- Modality: image + text
+- Language: Chinese (Simplified/Traditional text may appear in source inscriptions)
+- License: CC BY 4.0
+- Total images: 2331
+- Total question items (`questions/questions_all.jsonl`): 26137
+
+## Supported Tasks
+
+The release includes multiple task files in `questions/by_type/`:
+
+- `CC.jsonl` (922)
+- `ER_choice.jsonl` (2875)
+- `ER_fillin.jsonl` (1279)
+- `IR.jsonl` (2351)
+- `ITT_MHQA_choice.jsonl` (5290)
+- `ITT_MHQA_fillin.jsonl` (5290)
+- `MITT_MHQA_choice.jsonl` (330)
+- `MITT_MHQA_fillin.jsonl` (330)
+- `PR.jsonl` (922)
+- `SR.jsonl` (1792)
+- `TR_choice.jsonl` (1312)
+- `TR_fillin.jsonl` (1312)
+- `TTI.jsonl` (1210)
+- `VA.jsonl` (922)
+
+## Repository Structure
+
+- `questions/by_type/*.jsonl`: per-task split files
+- `images/**`: image assets used by questions
+- `kb/knowledge_base.json`: structured knowledge base
+- `annotations/**`: annotation resources
+
+## Data Format
+
+Each line in question files is a JSON object. Typical fields include:
+
+- `qid`: unique question id
+- `image_id`: image identifier
+- `phase`: benchmark phase
+- `question_no`: question index in phase
+- `type`: task type
+- `answer_format`: expected output format (`text` or `json`)
+- `prompt`: full question prompt with output constraints
+- `ground_truth`: reference answer (string or structured object)
+- `image_paths`: related image path list (relative paths)
+
+Example (single line from `questions/questions_all.jsonl`):
+
+```json
+{
+	"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。",
+	"ground_truth": "...",
+	"image_paths": ["images/IMG000002_0.jpg", "images/IMG000002_1.jpg"]
+}
+```
+
+## Usage
+
+### 1) Load by `datasets`
+
+```python
+from datasets import load_dataset
+
+# Replace with your dataset repo id
+repo_id = "g41/KnowCP"
+
+all_ds = load_dataset(repo_id, data_files={"train": "questions/questions_all.jsonl"})
+ir_ds = load_dataset(repo_id, data_files={"train": "questions/by_type/IR.jsonl"})
+
+print(all_ds["train"][0]["qid"])
+print(ir_ds["train"][0]["type"])
+```
+
+### 2) Run evaluation with this project (`github_repo/run_eval_for_hf.py`)
+
+[GitHub link](https://github.com/41-edu/KnowCP)，This runner expects local files. If your dataset is already on Hugging Face, download it first and point `--workspace-root` to the parent folder that contains both `github_repo` and `hf_repo`.
+
+```bash
+# Example local layout after download:
+# D:/mylib/benchmark/
+#   ├─ github_repo/
+#   └─ hf_repo/
+
+python github_repo/run_eval_for_hf.py \
+	--workspace-root D:/mylib/benchmark \
+	--items-dir D:/mylib/benchmark/hf_repo/questions/by_type \
+	--model bytedance-seed/seed-2.0-lite \
+	--api-base https://openrouter.ai/api/v1 \
+	--api-key YOUR_API_KEY
+```
+
+## Intended Uses
+
+- Benchmarking multimodal models on Chinese painting understanding
+- Evaluating grounding, OCR-style extraction, and text reasoning
+- Multi-turn visual QA experiments
+
+## Limitations
+
+- Prompts are optimized for benchmark consistency, not for general chat behavior.
+- Some tasks require strict JSON output formatting.
+- Evaluation metrics and threshold settings may differ by task type.
+
+## Citation
+
+If you use this dataset, please cite the project repository and dataset page:
+
+- Code: https://github.com/41-edu/KnowCP
+- Dataset: https://huggingface.co/datasets/g41/KnowCP