stevafernandes/Fine-Tuning-HunyuanOCR

MuseumSCAT Label Transcription

HunyuanOCR-1.5 fine-tuned to transcribe collection dates (verbatimDate) and localities (verbatimLocality) from Natural History Museum of Denmark specimen label images, with per-field confidence scores.

Deployment

1. 1.Create a Gradio Space and upload the contents of this directory.
2. 2.Add the fine-tuned LoRA adapter either by:
3. 3.copying the adapter directory (from finetune.py --output-dir) into the Space as adapter/, or
4. 4.setting the ADAPTER_REPO Space variable to a Hugging Face model repo containing the adapter.
5. 5.Without an adapter, the app runs the base tencent/HunyuanOCR model.

GPU hardware (or ZeroGPU) is recommended; CPU inference works but takes minutes per image.