graziasveva93/epo-examiner-Qwen3.5-27B
VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer Open Weights Cold
The graziasveva93/epo-examiner-Qwen3.5-27B model is a 27 billion parameter Qwen3.5 architecture, fine-tuned for EPO patentability classification. It is a student model, distilled from a stronger reasoning model, designed to reason like patent examiners. This model excels in specialized patent analysis tasks, leveraging its native blocks for structured reasoning.
Loading preview...
Model Overview
The graziasveva93/epo-examiner-Qwen3.5-27B is a specialized 27 billion parameter language model built on the Qwen3.5 architecture. It is an instruction-tuned student model, specifically designed for EPO patentability classification. This model was developed as part of the research detailed in "Teaching Large Language Models to Reason Like Patent Examiners."
Key Capabilities
- Patentability Classification: Fine-tuned to perform classification tasks related to European Patent Office (EPO) patentability.
- Reasoning Distillation: Inherits reasoning capabilities from its base model,
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled, which was itself distilled from Claude 4.6 Opus. - Structured Reasoning: Utilizes native
<think>...</think>blocks, enabling it to generate structured reasoning traces, similar to how patent examiners approach problems. - LoRA Adapter: Employs a LoRA adapter with
r=64andalpha=64targeting key attention and feed-forward modules for efficient fine-tuning.
Good For
- Automated Patent Analysis: Ideal for applications requiring automated assessment or classification of patent documents based on EPO guidelines.
- Research in AI for Legal/IP: Useful for researchers exploring the application of LLMs in intellectual property and legal reasoning.
- Developing Reasoning-Grounded AI: Provides a foundation for building systems that require transparent, step-by-step reasoning in specialized domains.