README.md

metadata

title: GROBID
emoji: 📚
colorFrom: purple
colorTo: green
sdk: docker
app_port: 8070
pinned: false
license: apache-2.0
fullWidth: true
short_description: GROBID — scholarly PDF to structured TEI XML

GROBID on a Hugging Face Docker Space

GROBID (GeneRation Of BIbliographic Data) turns scholarly PDFs into structured TEI XML — header (title/authors/affiliations/abstract), body full text, and parsed bibliography.

Deployed for the psychoguineapig.com box, which retired GROBID locally on 2026-10-09 because the JVM was the heaviest memory consumer on the host (1.26 GiB RAM + 1.63 GiB swap) and nothing was calling it. Here it runs on Hugging Face free CPU compute instead — no RAM on our box, no GPU quota spent.

Image choice

grobid/grobid:0.9.1-crf — the CRF-only image, 0.51 GB on disk. The -full image is 14.8 GB and is documented as wanting a GPU for reasonable runtime; free CPU Basic Spaces have no GPU (2 vCPU / 16 GB RAM), so CRF is the correct trade: it is the recommended image for fulltext/RAG pre-processing.

Endpoints

• https://strangerinastrangeland-grobid.hf.space/api/isalive — health
• /api/processFulltextDocument (multipart PDF) — header + body + references
• /api/processHeaderDocument — metadata only (cheapest, ~2 GB RAM)
• /api/processReferences — bibliography only

Interactive console: https://strangerinastrangeland-grobid.hf.space/

Notes

• Recipe mirrors the long-standing kermitt2/grobid Space: vendor image, root, writable grobid-home/tmp, explicit grobid-service CMD. ~20 GROBID Spaces exist on the Hub.
• HF Spaces cannot set --init / --ulimit core=0 (which GROBID's C++ PDF parser normally wants). If the container dies on a malformed PDF, that is the first suspect.
• Free CPU Basic Spaces sleep after 48 h idle; the next request wakes them (~30 s cold start).