Commit History
28 days ago
-
Publish the 2026-10-06 segmentation models, in both engines
Daniel EcerSegmentation decides what a region of a document is before any other model sees it, so a mistake there is one no later model can recover from. These are the first retrained segmentation models here: GROBID 0.9.0's corpus with training data generated from 20 SciELO Preprints and Open Research Europe articles added to it, trained in both engines on the same inputs in the same order so a difference between them is the engine rather than the data. They predict one label GROBID 0.9.0's segmentation model does not, `<review>`, for the peer reviews a preprint server publishes beside the article. The README says so where it is hard to miss: a consumer that swaps one of these in without handling it is given a region it has no case for. `training.yml` records what each run read, what it was given and what it reported — including that the wapiti run reached its 2000-iteration cap rather than its stop criterion, and that delft early-stopped at epoch 381 on a best epoch of 331. The generated half…08e5e4a
-
Add the 2026-09-18 delft reference segmenter and citation models
Verified signaturedd39a9d
-
Publish the 2026-09-08 reference-segmenter and citation models, with their grobid-only controls
Verified signaturee266cc5