Add BERTimbau NER Gradio app for pt-ner

README.md

@@ -1,13 +1,50 @@
----
-title: Pt Ner Bertimbau
-emoji: 🐢
-colorFrom: blue
-colorTo: green
-sdk: gradio
-sdk_version: 6.24.0
-python_version: '3.12'
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: pt-ner BERTimbau NER
+emoji: 🚀
+colorFrom: blue
+colorTo: green
+sdk: gradio
+sdk_version: 5.50.0
+app_file: app.py
+short_description: Portuguese NER API for the pt-ner application
+python_version: "3.12"
+startup_duration_timeout: 30m
+---
+
+# pt-ner BERTimbau NER Space
+
+Public HTTP API backing [pt-ner](https://github.com/sillohq/starter): Portuguese named-entity recognition with [`marquesafonso/bertimbau-large-ner-total`](https://huggingface.co/marquesafonso/bertimbau-large-ner-total).
+
+## Deploy
+
+Create the Space (ZeroGPU, public):
+
+```bash
+hf repos create YOUR_USER/pt-ner-bertimbau --type space --space-sdk gradio \
+  --flavor zero-a10g --public --exist-ok
+```
+
+Upload this directory:
+
+```bash
+hf upload space/. --repo-type space --repo YOUR_USER/pt-ner-bertimbau \
+  --exclude "**/__pycache__/**"
+```
+
+When the Space is running, set in the pt-ner app (Docker `.env`):
+
+```bash
+NER_BACKEND=space
+NER_SPACE_URL=https://YOUR_USER-pt-ner-bertimbau.hf.space
+```
+
+## API
+
+Gradio exposes `predict(text) -> list[entity]`. Each entity has:
+
+- `label` — entity type
+- `text` — surface form
+- `start` / `end` — character offsets in the input text
+- `score` — model confidence
+
+The pt-ner application calls this endpoint over HTTP; it does not load transformers locally.

app.py

@@ -0,0 +1,68 @@
+"""Gradio ZeroGPU Space — Portuguese NER API for pt-ner."""
+
+from __future__ import annotations
+
+import spaces
+import gradio as gr
+import torch
+from transformers import pipeline
+
+MODEL_ID = "marquesafonso/bertimbau-large-ner-total"
+
+ner = pipeline(
+    "ner",
+    model=MODEL_ID,
+    aggregation_strategy="simple",
+    device=0,
+    torch_dtype=torch.bfloat16,
+)
+ner.model.to("cuda")
+
+
+def _serialize(raw: list[dict]) -> list[dict]:
+    entities: list[dict] = []
+    for item in raw:
+        start = int(item["start"])
+        end = int(item["end"])
+        entities.append(
+            {
+                "label": str(item.get("entity_group") or item.get("entity") or "MISC"),
+                "text": str(item.get("word") or ""),
+                "start": start,
+                "end": end,
+                "score": float(item["score"]) if item.get("score") is not None else None,
+            }
+        )
+    return entities
+
+
+@spaces.GPU(duration=120)
+def predict(text: str) -> list[dict]:
+    """Extract named entities from Portuguese text with character offsets."""
+    cleaned = (text or "").strip()
+    if not cleaned:
+        return []
+    return _serialize(ner(cleaned))
+
+
+EXAMPLES = [
+    ["João Silva mora em Lisboa."],
+    ["Maria Santos trabalha no Banco de Portugal, no Porto."],
+]
+
+demo = gr.Interface(
+    fn=predict,
+    inputs=gr.Textbox(label="Texto", lines=6, placeholder="Cole texto em português…"),
+    outputs=gr.JSON(label="Entidades"),
+    title="BERTimbau NER — pt-ner",
+    description=(
+        "API Space for the pt-ner application. "
+        f"Model: [{MODEL_ID}](https://huggingface.co/{MODEL_ID})."
+    ),
+    examples=EXAMPLES,
+    cache_examples=True,
+    cache_mode="lazy",
+)
+
+if __name__ == "__main__":
+    demo.launch(mcp_server=True)

requirements.txt

@@ -0,0 +1,3 @@
+transformers
+accelerate
+sentencepiece