Update app.py

app.py

@@ -7,412 +7,422 @@ from transformers import (
     AutoModelForSequenceClassification
 )
 
+from peft import PeftModel
 
-# ============================================================
-# GOSHAWK INSURANCE AI
-# DistilBERT - Insurance Classification
-# CPU OPTIMIZED
-# ============================================================
 
-MODEL_PATH = "./"
-DEVICE = torch.device("cpu")
+# ================================================================
+# CONFIG
+# ================================================================
 
-MAX_LENGTH = 256
+BASE_MODEL = "distilbert-base-uncased"
+ADAPTER_PATH = "."
 
+MAX_LENGTH = 128
 
-# ============================================================
-# MODEL YÜKLEME
-# ============================================================
+LABELS = {
+    0: "negative",
+    1: "neutral",
+    2: "positive"
+}
 
-print("=" * 60)
-print("GOSHAWK INSURANCE AI")
-print("Model yükleniyor...")
-print("=" * 60)
-
-try:
 
-    tokenizer = AutoTokenizer.from_pretrained(
-        MODEL_PATH,
-        local_files_only=True
-    )
-
-    model = AutoModelForSequenceClassification.from_pretrained(
-        MODEL_PATH,
-        local_files_only=True
-    )
+# ================================================================
+# DEVICE
+# ================================================================
 
-    model.to(DEVICE)
-    model.eval()
-
-except Exception as error:
-
-    print("=" * 60)
-    print("MODEL YÜKLEME HATASI")
-    print(error)
-    print("=" * 60)
-
-    raise error
-
-
-# ============================================================
-# MODEL BİLGİLERİ
-# ============================================================
+DEVICE = torch.device(
+    "cuda" if torch.cuda.is_available() else "cpu"
+)
 
 print("=" * 60)
-print("MODEL BAŞARIYLA YÜKLENDİ")
+print("INSURANCE SENTIMENT AI")
 print("=" * 60)
 
 print("Device:", DEVICE)
-print("Model:", type(model).__name__)
-print("Num labels:", model.config.num_labels)
 
-print("ID2LABEL:")
-print(model.config.id2label)
+if torch.cuda.is_available():
+    print("GPU:", torch.cuda.get_device_name(0))
 
-print("LABEL2ID:")
-print(model.config.label2id)
 
-print("=" * 60)
+# ================================================================
+# CHECK FILES
+# ================================================================
 
+print("\nChecking model files...")
 
-# ============================================================
-# LABEL OKUMA
-# ============================================================
+required_files = [
+    "adapter_config.json",
+    "adapter_model.safetensors",
+    "tokenizer.json",
+    "tokenizer_config.json"
+]
 
-def get_label(index):
+for file in required_files:
 
-    label = model.config.id2label.get(
-        index,
-        f"LABEL_{index}"
+    path = os.path.join(
+        ADAPTER_PATH,
+        file
     )
 
-    return str(label)
-
-
-# ============================================================
-# METİN ANALİZİ
-# ============================================================
-
-def predict(text):
-
-    # Boş metin kontrolü
-    if text is None or not text.strip():
-
-        return {
-            "Durum": "HATA",
-            "Mesaj": "Lütfen bir sigorta metni girin."
-        }
-
-
-    text = text.strip()
-
-
-    try:
+    print(
+        file,
+        "->",
+        "OK" if os.path.exists(path) else "MISSING"
+    )
 
-        # ----------------------------------------------------
-        # TOKENIZER
-        # ----------------------------------------------------
 
-        inputs = tokenizer(
-            text,
-            return_tensors="pt",
-            truncation=True,
-            padding=True,
-            max_length=MAX_LENGTH
-        )
+# ================================================================
+# TOKENIZER
+# ================================================================
 
+print("\nLoading tokenizer...")
 
-        # ----------------------------------------------------
-        # CPU
-        # ----------------------------------------------------
+tokenizer = AutoTokenizer.from_pretrained(
+    ADAPTER_PATH,
+    use_fast=True
+)
 
-        inputs = {
-            key: value.to(DEVICE)
-            for key, value in inputs.items()
-        }
+print("Tokenizer loaded.")
 
 
-        # ----------------------------------------------------
-        # MODEL
-        # ----------------------------------------------------
+# ================================================================
+# BASE MODEL
+# ================================================================
 
-        with torch.inference_mode():
+print("\nLoading base DistilBERT...")
 
-            outputs = model(**inputs)
+base_model = AutoModelForSequenceClassification.from_pretrained(
+    BASE_MODEL,
+    num_labels=3,
+    id2label=LABELS,
+    label2id={
+        "negative": 0,
+        "neutral": 1,
+        "positive": 2
+},
+    ignore_mismatched_sizes=True
+)
 
-            logits = outputs.logits[0]
+print("Base model loaded.")
 
-            probabilities = torch.softmax(
-                logits,
-                dim=-1
-            )
 
+# ================================================================
+# LOAD LoRA
+# ================================================================
 
-        # ----------------------------------------------------
-        # TÜM SINIFLAR
-        # ----------------------------------------------------
+print("\nLoading LoRA adapter...")
 
-        all_results = {}
+model = PeftModel.from_pretrained(
+    base_model,
+    ADAPTER_PATH
+)
 
-        for index in range(model.config.num_labels):
+print("LoRA adapter loaded.")
 
-            label = get_label(index)
 
-            probability = float(
-                probabilities[index]
-            ) * 100
+# ================================================================
+# MERGE LoRA
+# ================================================================
 
-            all_results[label] = round(
-                probability,
-                2
-            )
+print("\nMerging LoRA adapter...")
 
+try:
 
-        # ----------------------------------------------------
-        # EN YÜKSEK SINIF
-        # ----------------------------------------------------
+    model = model.merge_and_unload()
 
-        best_index = int(
-            torch.argmax(probabilities)
-        )
+    print("LoRA merged successfully.")
 
-        best_label = get_label(best_index)
+except Exception as e:
 
-        best_probability = (
-            float(probabilities[best_index]) * 100
-        )
+    print("Merge skipped:", str(e))
 
 
-        # ----------------------------------------------------
-        # SIRALA
-        # ----------------------------------------------------
+# ================================================================
+# DEVICE
+# ================================================================
 
-        sorted_results = dict(
-            sorted(
-                all_results.items(),
-                key=lambda item: item[1],
-                reverse=True
-            )
-        )
+model = model.to(DEVICE)
 
+model.eval()
 
-        # ----------------------------------------------------
-        # GÜVEN DURUMU
-        # ----------------------------------------------------
+print("\nModel ready.")
 
-        if best_probability >= 90:
 
-            confidence_status = "Çok yüksek"
+# ================================================================
+# PREDICTION
+# ================================================================
 
-        elif best_probability >= 70:
+def predict_sentiment(text):
 
-            confidence_status = "Yüksek"
-
-        elif best_probability >= 50:
-
-            confidence_status = "Orta"
-
-        else:
+    if text is None:
+        return {
+            "negative": 0.0,
+            "neutral": 0.0,
+            "positive": 0.0
+        }
 
-            confidence_status = "Düşük"
+    text = str(text).strip()
 
+    if not text:
 
-        # ----------------------------------------------------
-        # NEUTRAL KONTROLÜ
-        # ----------------------------------------------------
+        return {
+            "negative": 0.0,
+            "neutral": 0.0,
+            "positive": 0.0
+        }
 
-        neutral_warning = ""
+    # ------------------------------------------------------------
+    # Tokenize
+    # ------------------------------------------------------------
 
-        if (
-            "neutral" in best_label.lower()
-            and best_probability >= 90
-        ):
+    inputs = tokenizer(
+        text,
+        return_tensors="pt",
+        truncation=True,
+        max_length=MAX_LENGTH,
+        padding=True
+    )
 
-            neutral_warning = (
-                "Model Neutral sınıfına çok yüksek güven "
-                "veriyor. Farklı metinlerde de aynı sonuç "
-                "oluşuyorsa eğitim/label dağılımı kontrol "
-                "edilmelidir."
-            )
+    inputs = {
+        key: value.to(DEVICE)
+        for key, value in inputs.items()
+    }
 
+    # ------------------------------------------------------------
+    # Inference
+    # ------------------------------------------------------------
 
-        # ----------------------------------------------------
-        # SONUÇ
-        # ----------------------------------------------------
+    with torch.inference_mode():
 
-        return {
+        outputs = model(
+            **inputs
+        )
 
-            "Tahmin": best_label,
+        probabilities = torch.softmax(
+            outputs.logits,
+            dim=-1
+        )[0]
 
-            "Güven": (
-                f"{best_probability:.2f}%"
-            ),
+    # ------------------------------------------------------------
+    # Results
+    # ------------------------------------------------------------
 
-            "Güven Durumu": confidence_status,
+    result = {
+        "negative": float(
+            probabilities[0].item()
+        ),
 
-            "Tüm Sınıflar": {
-                label: f"{score:.2f}%"
-                for label, score
-                in sorted_results.items()
-            },
+        "neutral": float(
+            probabilities[1].item()
+        ),
 
-            "Neutral Kontrolü": neutral_warning,
+        "positive": float(
+            probabilities[2].item()
+        )
+    }
 
-            "Model Bilgisi": {
+    return result
 
-                "Model": type(model).__name__,
 
-                "Num Labels": model.config.num_labels,
+# ================================================================
+# DETAILED PREDICTION
+# ================================================================
 
-                "Device": str(DEVICE),
+def analyze_news(text):
 
-                "Max Length": MAX_LENGTH
+    if text is None or not str(text).strip():
 
-            }
+        return (
+            "Please enter an insurance news headline.",
+            "",
+            ""
+        )
 
-        }
+    text = str(text).strip()
 
+    probabilities = predict_sentiment(text)
 
-    except Exception as error:
+    predicted_label = max(
+        probabilities,
+        key=probabilities.get
+    )
 
-        return {
+    confidence = (
+        probabilities[predicted_label] * 100
+    )
 
-            "Durum": "MODEL HATASI",
+    result_text = (
+        f"Prediction: {predicted_label.upper()}\n"
+        f"Confidence: {confidence:.2f}%"
+    )
 
-            "Hata": str(error)
+    probability_text = (
+        f"Negative: "
+        f"{probabilities['negative'] * 100:.2f}%\n"
+        f"Neutral: "
+        f"{probabilities['neutral'] * 100:.2f}%\n"
+        f"Positive: "
+        f"{probabilities['positive'] * 100:.2f}%"
+    )
 
-        }
+    return (
+        result_text,
+        probability_text,
+        text
+    )
 
 
-# ============================================================
-# TEST ÖRNEKLERİ
-# ============================================================
+# ================================================================
+# EXAMPLES
+# ================================================================
 
 examples = [
 
     [
-        "Araç kazası sonucunda ön tampon ve far hasar gördü."
-    ],
-
-    [
-        "Sigortalı kasko poliçesi kapsamında hasar başvurusunda bulundu."
+        "Insurance companies reported strong financial results."
     ],
 
     [
-        "Trafik kazası nedeniyle araçta maddi hasar meydana geldi."
+        "The insurer suffered major losses after severe claims."
     ],
 
     [
-        "Poliçe kapsamında sağlık giderlerinin karşılanması talep edildi."
+        "The insurance company maintained stable operations."
     ],
 
     [
-        "Sigorta şirketi hasar dosyasını inceleyerek ödeme yaptı."
+        "The insurer reported record profits and strong growth."
     ],
 
     [
-        "Sigortalının poliçe teminatı kapsamında tazminat talebi bulunmaktadır."
+        "Insurance claims increased significantly after the disaster."
     ]
-
 ]
 
 
-# ============================================================
-# GRADIO ARAYÜZÜ
-# ============================================================
+# ================================================================
+# GRADIO UI
+# ================================================================
+
+DESCRIPTION = """
+# 🛡️ Insurance Sentiment AI
+
+AI-powered sentiment analysis for **insurance news headlines**.
+
+### Classes
+
+🟢 **Positive** — favorable insurance/business news  
+⚪ **Neutral** — factual or balanced news  
+🔴 **Negative** — unfavorable insurance/business news
+
+Model: **DistilBERT + LoRA**
+
+Validation Macro F1: **78.06%**
+"""
+
 
 with gr.Blocks(
-    title="Goshawk Insurance AI"
+    title="Insurance Sentiment AI"
 ) as demo:
 
     gr.Markdown(
-        """
-        # 🛡️ Goshawk Insurance AI
-
-        **Sigorta sektörüne özel DistilBERT yapay zeka modeli**
-
-        Sigorta metnini girin ve modelin hangi sınıfa
-        ait olduğunu tahmin etmesini sağlayın.
-        """
+        DESCRIPTION
     )
 
+    gr.Markdown(
+        "Enter an English insurance news headline:"
+    )
 
     with gr.Row():
 
         with gr.Column():
 
             text_input = gr.Textbox(
-                label="Sigorta Metni",
+                label="Insurance News",
                 placeholder=(
-                    "Örneğin: "
-                    "Araç kazası nedeniyle ön tampon hasar gördü..."
+                    "Example: "
+                    "Insurance companies reported "
+                    "strong financial results."
                 ),
-                lines=8
+                lines=5
             )
 
-
-            with gr.Row():
-
-                analyze_button = gr.Button(
-                    "🔍 Analiz Et",
-                    variant="primary"
-                )
-
-                clear_button = gr.Button(
-                    "🗑️ Temizle"
-                )
-
+            analyze_button = gr.Button(
+                "🔍 Analyze Sentiment",
+                variant="primary"
+            )
 
         with gr.Column():
 
-            output = gr.JSON(
-                label="Model Sonucu"
+            prediction_output = gr.Textbox(
+                label="Prediction",
+                lines=3
             )
 
+            probability_output = gr.Textbox(
+                label="Probabilities",
+                lines=4
+            )
 
-    # --------------------------------------------------------
-    # ÖRNEKLER
-    # --------------------------------------------------------
+            analyzed_text = gr.Textbox(
+                label="Analyzed Headline",
+                lines=3
+            )
 
-    gr.Examples(
-        examples=examples,
+    analyze_button.click(
+        fn=analyze_news,
         inputs=text_input,
-        label="Örnek Sigorta Metinleri"
+        outputs=[
+            prediction_output,
+            probability_output,
+            analyzed_text
+        ]
     )
 
-
-    # --------------------------------------------------------
-    # BUTONLAR
-    # --------------------------------------------------------
-
-    analyze_button.click(
-        fn=predict,
+    text_input.submit(
+        fn=analyze_news,
         inputs=text_input,
-        outputs=output
+        outputs=[
+            prediction_output,
+            probability_output,
+            analyzed_text
+        ]
+    )
+
+    gr.Examples(
+        examples=examples,
+        inputs=text_input
     )
 
+    gr.Markdown(
+        """
+---
+
+### Model Information
 
-    clear_button.click(
-        fn=lambda: "",
-        inputs=None,
-        outputs=text_input
+- **Base:** DistilBERT
+- **Fine-tuning:** LoRA / PEFT
+- **Task:** Insurance News Sentiment Classification
+- **Labels:** Negative / Neutral / Positive
+- **Validation Accuracy:** 79.75%
+- **Validation Macro F1:** 78.06%
+"""
     )
 
 
-# ============================================================
-# SERVER
-# ============================================================
+# ================================================================
+# LAUNCH
+# ================================================================
 
 if __name__ == "__main__":
 
-    port = int(
-        os.environ.get(
-            "PORT",
-            7860
-        )
-    )
-
     demo.launch(
         server_name="0.0.0.0",
-        server_port=port
+        server_port=int(
+            os.environ.get(
+                "PORT",
+                7860
+            )
+        ),
+        show_error=True
     )