Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,93 @@
+"""
+DIPPER Paraphrase API — HuggingFace Spaces deployment.
+Exposes DIPPER-lite (T5-large, 1.0B) as a Gradio API for de-AI rewriting.
+"""
+import gradio as gr
+import torch
+from transformers import T5Tokenizer, T5ForConditionalGeneration
+import time
+
+print("[DIPPER] Loading model...")
+t0 = time.time()
+tokenizer = T5Tokenizer.from_pretrained("SamSJackson/paraphrase-dipper-no-ctx")
+model = T5ForConditionalGeneration.from_pretrained("SamSJackson/paraphrase-dipper-no-ctx")
+model.eval()
+print(f"[DIPPER] Model loaded in {time.time()-t0:.1f}s")
+
+def paraphrase(text: str, lex: int = 40, order: int = 40) -> str:
+    """
+    Paraphrase text using DIPPER.
+    
+    Args:
+        text: Input text (one paragraph recommended, max ~300 words)
+        lex: Lexical control (0=max change, 100=no change). Default 40.
+        order: Order control (0=max reorder, 100=no reorder). Default 40.
+    
+    Returns:
+        Paraphrased text.
+    """
+    if not text or not text.strip():
+        return ""
+    
+    text = text.strip()
+    prefix = f"lexical = {lex}, order = {order} "
+    input_text = prefix + text
+    
+    inputs = tokenizer(input_text, return_tensors="pt", max_length=512, truncation=True)
+    
+    t0 = time.time()
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            max_length=512,
+            do_sample=False,
+            num_beams=5,
+            num_return_sequences=1
+        )
+    elapsed = time.time() - t0
+    
+    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    print(f"[DIPPER] {len(text)} -> {len(result)} chars in {elapsed:.1f}s")
+    return result
+
+def batch_paraphrase(text: str, lex: int = 40, order: int = 40) -> str:
+    """
+    Paraphrase multi-paragraph text by processing each paragraph separately.
+    
+    Args:
+        text: Input text (can be multiple paragraphs separated by blank lines)
+        lex: Lexical control (0=max change, 100=no change). Default 40.
+        order: Order control (0=max reorder, 100=no reorder). Default 40.
+    
+    Returns:
+        Paraphrased text with paragraphs rejoined.
+    """
+    if not text or not text.strip():
+        return ""
+    
+    paragraphs = [p.strip() for p in text.split('\n\n') if p.strip()]
+    results = []
+    
+    for i, para in enumerate(paragraphs):
+        print(f"[DIPPER] Paragraph {i+1}/{len(paragraphs)}")
+        result = paraphrase(para, lex=lex, order=order)
+        results.append(result)
+    
+    return '\n\n'.join(results)
+
+# Gradio interface
+with gr.Blocks(title="DIPPER Paraphrase API") as demo:
+    gr.Markdown("# DIPPER Paraphrase API\nRewrite text to bypass AI detection. Use the API endpoint for programmatic access.")
+    
+    with gr.Row():
+        with gr.Column():
+            input_text = gr.Textbox(label="Input Text", lines=10, placeholder="Paste text here...")
+            lex_slider = gr.Slider(0, 100, value=40, step=10, label="Lexical Control (40=optimal)")
+            ord_slider = gr.Slider(0, 100, value=40, step=10, label="Order Control (40=optimal)")
+            btn = gr.Button("Paraphrase", variant="primary")
+        with gr.Column():
+            output_text = gr.Textbox(label="Paraphrased Text", lines=10)
+    
+    btn.click(fn=batch_paraphrase, inputs=[input_text, lex_slider, ord_slider], outputs=output_text)
+
+demo.launch()