shadowfurynon/prescription_reader

Script+ Reader — handwritten Indian prescription recognition

TrOCR (LoRA-merged) + LayoutLMv3 field pairing + curated Indian brand→generic vocabulary + beam-consensus abstention gate. The model runs inside this Space — upload a prescription photo and get an explained, confidence-tiered reading.

Deploying (one-time, ~20 min)

Space repos on HF cap free storage at 1 GB — too small for a merged TrOCR checkpoint. So checkpoints live in a separate model repo and get pulled onto the Space's disk automatically the first time rx_pipeline.py runs.

1. Create the Space: huggingface.co/new-space → SDK: Gradio. Upload app.py, rx_pipeline.py, requirements.txt, README.md.
2. Create a separate model repo: huggingface.co/new → type Model, e.g. YOURNAME/scriptplus-checkpoints (private is fine).
3. Push your checkpoints there from Colab (Drive already mounted):

   !pip -q install huggingface_hub
   from huggingface_hub import HfApi
   api = HfApi(token="hf_...your write token...")
   api.upload_folder(folder_path="/content/drive/MyDrive/rx_system/m_htr_merged",
                     repo_id="YOURNAME/scriptplus-checkpoints", repo_type="model",
                     path_in_repo="m_htr_merged")
   api.upload_folder(folder_path="/content/drive/MyDrive/rx_system/m2_model",
                     repo_id="YOURNAME/scriptplus-checkpoints", repo_type="model",
                     path_in_repo="m2_model")
   api.upload_file(path_or_fileobj="/content/drive/MyDrive/rx_system/abstention_gate.joblib",
                   repo_id="YOURNAME/scriptplus-checkpoints", repo_type="model",
                   path_in_repo="abstention_gate.joblib")
   

4. On the Space page → Settings → Variables and secrets:
   • Add secret HF_TOKEN = a read-scope token (needed if your model repo is private).
   • Add variable RX_CKPT_REPO = YOURNAME/scriptplus-checkpoints (skip this if you kept the default shadowfurynon/scriptplus-checkpoints).
5. Restart the Space. On first boot rx_pipeline.py downloads the checkpoints from the model repo into local disk automatically — watch the Logs tab, this takes a few minutes the first time only.

Hardware

Free CPU tier works but a page takes ~2–5 min (8 beams × every crop is heavy). For demos, switch Space hardware to T4 small while presenting, then back to CPU.

Env knobs

• RX_EXPAND_VOCAB=0 → curated-only vocabulary (recall ceiling-limited)
• RX_GATE_TH → abstention gate threshold (default 0.5)
• RX_CKPT_REPO → HF model repo holding checkpoints (default shadowfurynon/scriptplus-checkpoints)
• HF_TOKEN → set as a Space secret if the checkpoint model repo is private