README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- text-classification
language:
- en
size_categories:
- 1K<n<10K
tags:
- patent
- embeddings
- benchmark
- text-embedding
---

# Patent Technology Classification (Bloom Taxonomy)

## Dataset Description

This dataset is part of **PatenTEB**, a comprehensive benchmark for evaluating text embedding models on patent-specific tasks. PatenTEB comprises 15 tasks across retrieval, classification, paraphrase detection, and clustering, with 2.06 million examples designed to reflect real-world patent analysis workflows.

**Paper**: [PatenTEB: A Comprehensive Benchmark and Model Family for Patent Text Embedding](https://arxiv.org/abs/2510.22264)

### Task Details

- **Task Name**: `class_bloom`
- **Task Type**: Classification
- **Test Samples**: 7,347
- **Training Samples**: 11,636 (20% training subset for evaluation)

Patent citation timing classification distinguishing patents by their temporal citation trajectory. 
Patents are classified into three categories:
- **Early bloom**: Patents ranking in the top decile for 5-year forward citations within their IPC3 domain (rapid impact)
- **Late bloom**: Patents ranking in the top decile for lifetime citations but bottom decile for 5-year citations (delayed recognition)  
- **Normal**: Patents outside these extremes (balanced class)

This formulation tests whether embeddings capture temporal citation dynamics and innovation impact patterns.

### Dataset Structure

This is a text classification task.

**Splits:**
- `test`: Full test set for evaluation
- `train`: 20% subset of training data used during evaluation

**Columns:**
- `q`
- `text`
- `label`

### Data Sample

Below is a 5-row preview of the test set:

```csv
q,text,label
000-303-150-320-965,semiconductor device and manufacture thereof [SEP] prevent formation depth and dispersion characteristics of a diffusion layer from being unstable....,early
000-806-432-248-978,manufacturing customized articles of sports apparel based on sensor data [SEP] provide articles of sports apparel customized for a person and a met...,early
002-198-800-516-400,capability expanding method and system for hand held device [SEP] provide a frame work enabling a hand held device to discriminate its peripheral d...,early
002-807-631-758-118,method for adjusting misalignment of data head and data detecting apparatus [SEP] speedily measure the error in the position of a detector and to p...,early
005-452-493-340-58X,single crystal pulling apparatus [SEP] obtain an apparatus for pulling up a single crystal capable of lessening the production of defective article...,normal
```

### Evaluation Metrics

This task uses **Macro-F1** as the primary metric, computed by averaging per-class F1 scores.
Models are evaluated using a 20% training subset with logistic regression classifiers trained on embeddings.

## Usage

### Load Dataset

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("datalyes/{task_name}")

# Access test split
test_data = dataset['test']

# Access 20% training subset (for classification tasks)
train_data = dataset['train']

```

### Use with Sentence Transformers

```python
from sentence_transformers import SentenceTransformer

# Load a patent-specialized model
model = SentenceTransformer("datalyes/patembed-base")

# Encode patent texts
embeddings = model.encode(test_data['text'])
```

### Integrate with MTEB

This dataset is designed to be integrated with the MTEB (Massive Text Embedding Benchmark) framework. Integration with MTEB is in progress and will be available once the corresponding pull requests are accepted.

## Benchmark Context

This dataset is part of a larger benchmark suite:

| Benchmark Component | Description |
|-------------------|-------------|
| **PatenTEB** | 15 tasks covering retrieval, classification, paraphrase, clustering |
| **Test Data (Released)** | 319,320 examples across all 15 tasks |
| **Training/Validation Data** | 1.74 million examples (planned for future release) |
| **Total Dataset Size** | 2.06 million annotated instances |

**Note**: Currently, only the test split is publicly available. Training and validation data release is planned for a future date.

**All 15 Tasks (NEW to MTEB)**:
- 3 classification tasks: Bloom timing, NLI directionality, IPC3 classification
- 2 clustering tasks: IPC-based, Inventor-based  
- 8 retrieval tasks: 3 symmetric (IN/MIXED/OUT domain) + 5 asymmetric (fragment-to-full)
- 2 paraphrase tasks: Problem and solution paraphrase detection

**MTEB Integration**: Upcoming (PR in progress)

## Citation

If you use this dataset, please cite our paper:

```bibtex
@misc{ayaou2025patentebcomprehensivebenchmarkmodel,
      title={PatenTEB: A Comprehensive Benchmark and Model Family for Patent Text Embedding}, 
      author={Iliass Ayaou and Denis Cavallucci},
      year={2025},
      eprint={2510.22264},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.22264}
}
```

## License

This dataset is released under the **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)** license.

- You are free to share and adapt the material
- You must give appropriate credit
- You may not use the material for commercial purposes
- If you remix, transform, or build upon the material, you must distribute your contributions under the same license

For full license details, see: https://creativecommons.org/licenses/by-nc-sa/4.0/

## Contact

- **Authors**: Iliass Ayaou, Denis Cavallucci
- **Institution**: ICUBE Laboratory, INSA Strasbourg
- **GitHub**: [github.com/iliass-y/patenteb](https://github.com/iliass-y/patenteb)
- **HuggingFace**: [huggingface.co/datalyes](https://huggingface.co/datalyes)