Skip to content

LLM use

Work in progress

This section needs further work and will be revised.

bibr selectively uses different approaches to deal with the diverse formatting of scientific papers and support richer extraction. We want to make accurate extraction practical with small, open-source models, while giving researchers control over how their papers are processed.

Development

bibr 🦫 was developed together with AI agentic tools and workflows. Still, it took me months to iteratively develop and manually validate as much of the pipeline as possible. Without these tools, building something of this scope on my own would probably have taken years—a real testament to the GROBID team.

Extraction

bibr combines native parsing, rules, specialized ML models, and selected LLM calls. LLMs help with tasks such as front-page metadata, where publishers arrange titles, authors, and affiliations differently.

On bibr's development papers, large models barely differ, while smaller ones lose accuracy: off the shelf, 4–8B models pass 6 to 11 points fewer papers than Qwen3.8-27B. Fine-tuning on bibr's own tasks closes about half of that gap, so bibr's local defaults are fine-tuned Qwen3.5 models: a 4B model 2.8 points behind the 27B and a 2B model 6.3 points behind. The hosted recommendation, Claude Haiku 5.5, is within a point of the 27B. These are development-set results; choosing an extraction model has the numbers and their limits.

  • OCR and extraction models can run locally or through cloud providers.
  • --no-llm disables downstream LLM extraction, with reduced output. PDF OCR is configured separately and may still use vision-language models or remote services.

See configuration for options, evaluation guidance for checking your own outputs, and known limitations for accuracy and language/domain coverage.