vineetsinghvats/SynthIPData

---
license: mit
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- patent
- legal
- synthetic-data
- office-actions
- USPTO
- rare-categories
- data-augmentation
size_categories:
- 1K<n<10K
---

# SynthIPData: Synthetic Patent Office Action Rejections for Rare Categories

## Dataset Description

SynthIPData is the first synthetic dataset targeting rare patent office action rejection categories. It contains 1,800 synthetic patent rejection documents generated by a LoRA-fine-tuned Mistral-7B model, trained on 6,161 real USPTO office actions.

### Why This Dataset?

Patent office action rejections are critical documents in intellectual property law. However, certain rejection-type x technology-area combinations are severely underrepresented in existing data. For example, 35 USC 101 rejections in Materials/Coatings (USPC 428) account for fewer than 100 documents in the entire 2020-2024 USPTO corpus. This data scarcity prevents AI systems from learning to handle these rare but important cases.

SynthIPData addresses this gap by generating high-fidelity synthetic office actions that are:
- **Indistinguishable from real text** (perplexity ratio 1.04 vs real documents)
- **Effective for retrieval augmentation** (+200% recall for rarest categories)
- **Useful in few-shot learning** (+12% F1 improvement when real data is scarce)

## 8 Rare Categories

| Category | Rejection Type | USPC Class | Technology Area | Real Seeds | Synthetic |
|---|---|---|---|---|---|
| 101_ai_ml | 35 USC 101 | 706 | AI/Neural Networks | 2,424 | 240 |
| 112_ai_ml | 35 USC 112 | 706 | AI/Neural Networks | 1,552 | 240 |
| dp_ai_ml | Double Patenting | 706 | AI/Neural Networks | 1,184 | 120 |
| 101_semiconductors | 35 USC 101 | 257 | Semiconductors | 363 | 240 |
| 101_surgical | 35 USC 101 | 606 | Surgical Instruments | 230 | 240 |
| 101_crypto | 35 USC 101 | 380 | Cryptography | 181 | 240 |
| 101_batteries | 35 USC 101 | 429 | Batteries/Fuel Cells | 152 | 240 |
| 101_materials | 35 USC 101 | 428 | Materials/Coatings | 75 | 240 |

## Data Sources

- **Real seeds**: USPTO Office Actions Weekly Archives (OACT), 2020-2024
- **Category discovery**: USPTO PTOFFACT dataset, 2014-2017 (2.4M office actions analyzed)
- **Synthetic generation**: Mistral-7B fine-tuned with LoRA on 5,544 real office actions

## Evaluation Results

### Text Quality (Perplexity)
Average perplexity ratio: **1.04** (1.0 = identical to real, lower is better)

| Category | Real PPL | Synthetic PPL | Ratio |
|---|---|---|---|
| 101_ai_ml | 6.1 | 6.1 | 0.99 |
| 101_semiconductors | 7.3 | 5.4 | 0.74 |
| 101_surgical | 6.3 | 5.8 | 0.93 |
| 112_ai_ml | 4.6 | 4.7 | 1.04 |

### Retrieval Improvement (Rare Categories)
| Category | Real Only Recall@1 | + SynthIPData | Improvement |
|---|---|---|---|
| 101_materials | 6.7% | 20.0% | +200% |
| 101_batteries | 30.0% | 43.3% | +44% |
| 101_semiconductors | 55.6% | 61.1% | +10% |
| dp_ai_ml | 67.1% | 73.0% | +9% |

### Few-Shot Classification
With only 30 real examples per category:
| Method | F1 Macro |
|---|---|
| Few-shot only | 0.307 |
| + Paraphrasing | 0.311 |
| **+ SynthIPData** | **0.344 (+12%)** |

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("vineetsinghvats/SynthIPData")

# Access synthetic documents
for doc in dataset['synthetic']:
print(doc['category'], doc['title'])
print(doc['text'][:200])

# Access real seed metadata
for doc in dataset['real_seeds']:
print(doc['category'], doc['title'])
```

## Model Training Details

- **Base model**: Mistral-7B-v0.1
- **Fine-tuning**: LoRA (r=16, alpha=32)
- **Training data**: 5,544 real office actions
- **Training loss**: 0.528 (validation: 0.561)
- **Memorization rate**: 8.3% before filtering, 0% after Qdrant-based deduplication

## Citation

```bibtex
@dataset{singh2026synthipdata,
title={SynthIPData: Synthetic Data Augmentation for Rare Patent Office Action Rejection Categories},
author={Singh, Vineet},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/datasets/vineetsinghvats/SynthIPData}
}
```

## License

MIT License

## Contact

Vineet Singh - [GitHub](https://github.com/vineetsingh-vs/synthipdata) | [LinkedIn](https://linkedin.com/in/vineetsingh44)