Orionfold/patent-strategist-v3-unsloth

Orionfold/patent-strategist-v3-unsloth: Offline Patent Reasoning

This model is a specialized 8 billion parameter language model, fine-tuned from deepseek-ai/DeepSeek-R1-0528-Qwen3-8B using Unsloth. It is specifically designed for offline patent prosecution reasoning on Spark-class hardware (e.g., NVIDIA DGX Spark with GB10 and 128 GB unified memory).

Key Capabilities

• Patent Prosecution Workflows: Addresses tasks such as claim construction (including Markush groups and doctrine of equivalents), MPEP-grounded office-action argument drafting, and prior-art relevance analysis.
• Offline Operation: Enables privileged client text to be processed securely without relying on hosted frontier APIs, crucial for law firms and IP teams.
• IRAC-shaped Reasoning: Distills chain-of-thought reasoning from its base model onto a synthetic patent-reasoning corpus to produce structured reasoning chains.
• Unsloth-trained BF16 Merged Weights: Provides transformers-format weights suitable for continued training in Unsloth's 4-bit QLoRA workflow or for inference paths outside llama.cpp.

Good For

• Patent Attorneys and IP Teams: Ideal for professionals requiring secure, offline analysis of patent-related documents.
• Edge Device Deployment: Suitable for deployment on Spark-class hardware or comparable edge devices where data privacy and local processing are paramount.
• Specific Use Cases: Excels in claim construction, MPEP-grounded office-action responses, prior-art relevance, non-obviousness reasoning, and patent-licensing scenario analysis.

Known Limitations

• Terminology Drift: Exhibits minor terminology drifts, such as "metes-and-times" instead of "metes and bounds," inherited from the synthetic corpus.
• Fabricated Citations: Contains instances of fabricated MPEP citations (e.g., §2163.05(s)) due to corpus-generator artifacts, though the balance of probe answers cite real MPEP sections.