Enriching Patent Claim Generation with European Patent Dataset
Fine-tuning large language models (LLMs) on high-quality European patent data from the EPD dataset significantly improves patent claim generation quality compared to models trained on other datasets or GPT-4o. This method enhances claims' completeness and coherence, allowing models to better handle jurisdictional differences and real-world drafting challenges. The findings suggest that using polished granted claims provides stable linguistic patterns for superior output, though standard RLHF methods may not yield the expected gains in this specific task.