PaddleOCR-VL-1.5-OCSR

This repository hosts a merged fine-tuned OCSR model based on PaddleOCR-VL-1.5.

The task is:

• input: molecular structure image
• output: canonical SMILES

Base Model

• Base model family: PaddleOCR-VL-1.5
• Fine-tuning method: LoRA SFT, then merged export

Training Objective

The fine-tuning line is restricted to a single output space:

• canonical SMILES only

Incompatible target spaces such as ssml_normed, chemfig, and LaTeX formula targets were excluded from the main line.

Current Local Evaluation

canonical_smiles_main_v1

• canonical exact accuracy: 32.86%
• token micro F1: 70.35%
• valid SMILES rate: 71.84%
• mean fingerprint Tanimoto: 0.6992

ocsr_realworld_mixed_eval_v1p1

• canonical exact accuracy: 33.77%
• token micro F1: 70.18%
• valid SMILES rate: 75.84%
• mean fingerprint Tanimoto: 0.6849

Source-level observations

• relatively strong on uob
• moderate on uspto
• weak on real_world
• very weak on decimer handwritten structures
• weak on edu_chemc

Intended Use

This model is intended for:

• OCSR research
• benchmarking
• pipeline prototyping
• weak-domain error analysis

It is not yet a state-of-the-art production OCSR model.

Limitations

• performance is highly uneven across data domains
• handwritten chemical structures remain a major failure mode
• real-world photos/scans are still weak
• exact benchmark performance is below strong public OCSR baselines such as MolNexTR / MolScribe on standard UOB and USPTO comparisons

Files

This repository includes:

• merged model weights
• tokenizer files
• processor/preprocessor files
• remote-code files required by the custom model class

Notes

This model card reflects the current V2-1 baseline state in the associated project workspace.