cser-detector — structflo-cser weights
2-class page detector (chemical_structure, compound_label) for chemistry documents. Architecture: D-FINE-L (HGNet-V2 backbone, transformers DFineForObjectDetection, Apache-2.0), initialised from ustc-community/dfine-large-coco (Apache-2.0, COCO-only pretraining) and trained clean-room on synthetic pages rendered with RDKit from ChEMBL SMILES, then fine-tuned on an internal annotated corpus. Single-file .safetensors checkpoint with the model config in its metadata; load with structflo.cser.inference.dfine.DFineDetector.from_file.
Versions v0.1–v0.4 of this repo were Ultralytics YOLO11l checkpoints and are retired (AGPL-3.0 lineage); use v1.0 or later with structflo-cser >= 1.0.
| field | value |
|---|---|
| latest version | v1.0 (HF tag weights-v1.0) |
| file | best.safetensors |
| sha256 | 29962b7115159d93b6b6d3f16b96769a6ffc34e3590e6697687a8e3425ad0bc7 |
| requires | structflo-cser>=1.0.0,<2.0.0 |
Usage
from structflo.cser.pipeline import ChemPipeline
pipeline = ChemPipeline() # resolves the latest registered weights
pairs = pipeline.process("page.png")
Pin a version with ChemPipeline(weights="v1.0").
Licence and provenance
Weights: Apache-2.0. Training data: synthetic pages rendered with RDKit from ChEMBL
(CC-BY-SA-3.0) SMILES plus an internal annotated corpus (not released). See the package's
THIRD_PARTY_NOTICES.md for upstream attributions.