Create app.py

app.py

@@ -0,0 +1,67 @@
+import os
+import torch
+import gradio as gr
+from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
+
+MODEL_ID = "MoritzLaurer/deberta-v3-large-zeroshot-v2.0"
+
+os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
+torch.set_num_threads(int(os.environ.get("OMP_NUM_THREADS", "2")))
+torch.set_num_interop_threads(1)
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
+model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
+model.eval()
+
+clf = pipeline(
+    task="zero-shot-classification",
+    model=model,
+    tokenizer=tokenizer,
+    device=-1
+)
+
+def run_zero_shot(text, labels, hypothesis_template, multi_label, top_k):
+    text = (text or "").strip()
+    labels = (labels or "").strip()
+    hypothesis_template = (hypothesis_template or "").strip() or "This text is about {}"
+
+    if not text:
+        return {"error": "Enter some text."}
+    candidate_labels = [x.strip() for x in labels.split(",") if x.strip()]
+    if not candidate_labels:
+        return {"error": "Enter at least 1 label (comma-separated)."}
+
+    with torch.inference_mode():
+        out = clf(
+            sequences=text,
+            candidate_labels=candidate_labels,
+            hypothesis_template=hypothesis_template,
+            multi_label=bool(multi_label)
+        )
+
+    pairs = list(zip(out["labels"], out["scores"]))
+    pairs.sort(key=lambda x: x[1], reverse=True)
+    pairs = pairs[: max(1, int(top_k))]
+
+    return {
+        "top": {"label": pairs[0][0], "confidence_pct": round(pairs[0][1] * 100, 2)},
+        "all": [{"label": k, "confidence_pct": round(v * 100, 2)} for k, v in pairs],
+        "raw": out
+    }
+
+demo = gr.Interface(
+    fn=run_zero_shot,
+    inputs=[
+        gr.Textbox(label="Text", lines=4, value="I am wahhhh"),
+        gr.Textbox(label="Candidate Labels (comma-separated)", value="sad, happy, angry, neutral"),
+        gr.Textbox(label="Hypothesis Template", value="This text is about {}"),
+        gr.Checkbox(label="Multi-label", value=False),
+        gr.Slider(label="Top-K to show", minimum=1, maximum=25, value=5, step=1),
+    ],
+    outputs=gr.JSON(label="Output"),
+    title="Zero-Shot Classification (DeBERTa v3 Large, MoritzLaurer)",
+    allow_flagging="never"
+)
+
+if __name__ == "__main__":
+    demo.launch()