- Jupyter Notebook 55.6%
- Python 44.4%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| figures | ||
| tasks | ||
| .gitignore | ||
| LICENSE | ||
| README.md | ||
Oracle Bone Inscriptions Multi-modal Dataset (OBIMD)
Dataset Overview
The Oracle Bone Inscriptions Multi-modal Dataset (OBIMD) is the first large-scale corpus providing pixel-aligned rubbings and facsimiles, character-level annotations, and sentence-level transcriptions and reading sequences for oracle bone inscription (OBI) research. This dataset enables multi-modal learning across visual, structural, and linguistic dimensions, laying the foundation for end-to-end OBI recognition and interpretation.
Hugging Face Mirror
This dataset is available on Hugging Face Hub:🤗OBIMD Dataset
Dataset Statistics
- Total OBI images: 10,077 across five Shang Dynasty phases
- Annotated characters: 93,652
- Missing character positions: 21,667 (due to fragmentation)
- Sentences: 21,941 syntactically validated
- Non-sentential elements: 4,192
Data Structure
The dataset follows a three-level hierarchy:
- Image-level: Contains rubbing and facsimile pairs
- Sentence-level: Groups characters into meaningful units
- Character-level: Detailed annotations for each character

Data Fields
| Field Name | Description | Example |
|---|---|---|
| Facsimile | Path to facsimile image | facsimile/h00002.jpg |
| Rubbing | Path to rubbing image | rubbing/h00002.jpg |
| RubbingName | Short identifier | H2 |
| GroupCategory | Sentence type | InscriptionSentence1 |
| Position | Bounding box (x,y,w,h) | 558,581,80,218 |
| OrderNumber | Character order in sentence | 5 |
| Label | Main character label | xkubtjk815 |
| SubLabel | Secondary label | xkubtjk815 |
| SeatFont | Placeholder indicator | 0 |
| Mark | Special marker | -1 |
Data Modalities
- Rubbing images: Original scanned images of oracle bones
- Facsimile images: Expert reconstructions aligned with rubbings
- Character annotations: Bounding boxes and classifications
- Sentence groupings: Semantic organization of characters
- Reading sequences: Order of characters within sentences

Hierarchical Character Metadata Supplement
This folder contains supporting materials for the character hierarchy used in OBIMD. It includes:
- Sub-character Images: Images of sub-character variants, stored in a two-level folder structure. The top-level directory is named by the main character UID, and each subfolder is named by the corresponding sub-character UID.
- Sub-character to Glyph Code Point Mapping (XLSX): Maps each sub-character UID to the platform-specific glyph code point used by the Oracular Digital Platform font.
- Sub-character to Main-character Mapping (XLSX): Maps each sub-character UID to its parent main character UID in the Oracular Digital Platform.
- Main-character → Reference Modern Character (JSON): For each main-character UID, the JSON provides a platform-provided reference modern Chinese character (a standardized surrogate used for lookup and navigation). Note: this mapping is for reference only and is not intended as a final philological interpretation, since oracle bone inscriptions decipherment remains ongoing.
Usage Notes
The dataset is designed to support various research tasks:
- Character detection and recognition
- Sentence-level character clustering
- Character reordering within sentences
- Multi-modal analysis (rubbing vs. facsimile)
Technical Validation
Environment Prepare:
conda create --name obimd python==3.12.12
conda activate obimd
pip install -r tasks/requirements.txt
We provide baseline implementations and demonstration scripts for all three technical validation tasks for the OBIMD dataset, and the code is in here .
character-level detection
We benchmark oracle bone character detection and recognition using a four-way setting across image types (rubbing vs. facsimile) and character types (main vs. sub-character). Models are trained with YOLOv11l and evaluated on mAP, precision, recall, and F1 metrics. Results confirm the dataset’s effectiveness while also revealing difficulties in detecting blurred strokes and handling highly varied character shapes.
sentence-level clustering
This task groups characters into sentence-level clusters based on category and layout. We evaluate using standard clustering metrics such as AMI, NMI, ARI, and Purity. The results demonstrate promising intra-cluster consistency, while inter-sentence confusion suggests room for improvement in complex layouts.
character-level reordering
We assess sequential modeling of oracle bone inscriptions by recovering original reading order in shuffled sentences. A Transformer-based model predicts positional indices, achieving over 75% top-1 accuracy on main characters. These results show the feasibility of structure-aware ordering despite script irregularities.
Citation
If you use this dataset in your research, please cite: