Pilot-scale fine-tune of Llama-3.1-8B-Instruct for US utility patent claim drafting: given a detailed description, generate the full claim set (and the reverse: specification from a brief summary).

Training: 45,000 examples uniformly sampled from 2014–2023 US grants (PatentsView bulk data, CC-BY — attribution to PatentsView), 8K context, bf16 LoRA r=64, one epoch, effective batch 16. Final eval loss 0.8224.

Contamination statement: trained exclusively on pre-2024 grants; evaluate on 2024+ grants only.

Scope and limits: pilot for pipeline validation — keep input descriptions under ~6,000 tokens; dependent-claim completeness is the known weak suit. Larger paper-grade models (Qwen-14B, Gemma-4-31B, 16K context, full 616K-example corpus) are trained separately. Not legal advice; outputs require attorney review.

Prompt format: system prompt "You are an experienced patent attorney drafting US utility patent documents..." then user: "Draft the complete claim set for a patent application based on the following detailed description of the invention." followed by "=== DETAILED DESCRIPTION ===" and the description text.

Citation

bibtex

@misc{MikeTeodorescu_2026patentllmclaims,
  author       = {Mike Teodorescu},
  title        = {Llama-patentLLMClaimsSmall: A Llama-3.1-8B Pilot Model
                  for US Patent Claim Generation},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/miketeod/Llama-patentLLMClaimsSmall}},
  note         = {Fine-tuned on 45,000 US utility patent grants (2014--2023)
                  derived from PatentsView bulk data; 8K context, LoRA r=64}
}

Please also cite the base model (Grattafiori et al., 2024, The Llama 3 Herd of Models, arXiv:2407.21783) and the data source (PatentsView bulk data, CC-BY 4.0, patentsview.org).