Update model loading to prioritize local trained models for improved accuracy; adjust configuration for model paths and fallback options

dependencies/scoring_utils.py

@@ -184,31 +184,44 @@ def validate_input_file(input_path: Path, required_columns: list) -> pd.DataFram
 def load_polarity_model(model_variant: str, base_dir: Path, device: str = "cuda"):
     """
     Factory function to load polarity model by variant name.
-    
+
     Supported variants:
-      - "scibert": scibert/scibert_polarity/final_model
-      - "deberta": alternative_polarity/deberta/final_model
-      - "scideberta": alternative_polarity/scideberta/final_model
-    
+      - "scibert": scibert/scibert_polarity/final_model (F1=0.724 baseline)
+      - "deberta": training/outputs/deberta_polarity/final_model (F1=0.764, +5.5% - RECOMMENDED)
+      - "deberta_v3_small": training/outputs/deberta_v3_small_polarity/final_model (F1=0.754)
+      - "modernbert": training/outputs/modernbert_polarity/final_model (F1=0.741)
+      - "scideberta": training/outputs/scideberta_polarity/final_model (F1=0.737)
+
     Args:
         model_variant: Name of model variant
         base_dir: Base directory of project
         device: Device to load onto
-        
+
     Returns:
         Tuple of (tokenizer, model, device_obj)
-        
+
     Raises:
         ValueError: If model_variant not supported
         FileNotFoundError: If model directory doesn't exist
     """
+    # Feb 2026: New trained models from training/outputs/ (standardized comparison)
     variant_map = {
-        "scibert": base_dir / "scibert" / "scibert_polarity" / "final_model",
-        "deberta": base_dir / "alternative_polarity" / "deberta" / "deberta_v3_base_polarity_final_model",
-        "scideberta": base_dir / "alternative_polarity" / "scideberta" / "scideberta_full_polarity_final_model",
+        "scibert": base_dir / "training" / "outputs" / "scibert_polarity" / "final_model",
+        "deberta": base_dir / "training" / "outputs" / "deberta_polarity" / "final_model",  # BEST: F1=0.764
+        "deberta_v3_small": base_dir / "training" / "outputs" / "deberta_v3_small_polarity" / "final_model",
+        "modernbert": base_dir / "training" / "outputs" / "modernbert_polarity" / "final_model",
+        "scideberta": base_dir / "training" / "outputs" / "scideberta_polarity" / "final_model",
+        # Legacy models (pre-Feb 2026, kept for backwards compatibility)
+        "scibert_legacy": base_dir / "scibert" / "scibert_polarity" / "final_model",
+        "deberta_legacy": base_dir / "alternative_polarity" / "deberta" / "deberta_v3_base_polarity_final_model",
+        "scideberta_legacy": base_dir / "alternative_polarity" / "scideberta" / "scideberta_full_polarity_final_model",
     }
     hub_fallback_map = {
         "scibert": "Sina1138/Scibert_polarity_Review",
+        "scideberta": "KISTI-AI/Scideberta-full",  # Needs fine-tuning
+        "modernbert": "answerdotai/ModernBERT-base",  # Needs fine-tuning
+        "deberta": "microsoft/deberta-v3-base",  # Needs fine-tuning
+        "deberta_v3_small": "microsoft/deberta-v3-small",  # Needs fine-tuning
     }
 
     if model_variant not in variant_map:
@@ -224,31 +237,44 @@ def load_polarity_model(model_variant: str, base_dir: Path, device: str = "cuda"
 def load_topic_model(model_variant: str, base_dir: Path, device: str = "cuda"):
     """
     Factory function to load topic model by variant name.
-    
+
     Supported variants:
-      - "scibert": scibert/scibert_topic/final_model
-      - "deberta": alternative_topic/deberta/final_model
-      - "scideberta": alternative_topic/scideberta/final_model
-    
+      - "scideberta": training/outputs/scideberta_topic/final_model (F1=0.478 - BEST, RECOMMENDED)
+      - "deberta": training/outputs/deberta_topic/final_model (F1=0.450)
+      - "scibert": training/outputs/scibert_topic/final_model (F1=0.442)
+      - "deberta_v3_small": training/outputs/deberta_v3_small_topic/final_model (F1=0.381)
+      - "modernbert": training/outputs/modernbert_topic/final_model (F1=0.376)
+
     Args:
         model_variant: Name of model variant
         base_dir: Base directory of project
         device: Device to load onto
-        
+
     Returns:
         Tuple of (tokenizer, model, device_obj)
-        
+
     Raises:
         ValueError: If model_variant not supported
         FileNotFoundError: If model directory doesn't exist
     """
+    # Feb 2026: New trained models from training/outputs/ (standardized comparison)
     variant_map = {
-        "scibert": base_dir / "scibert" / "scibert_topic" / "final_model",
-        "deberta": base_dir / "alternative_topic" / "deberta" / "final_model",
-        "scideberta": base_dir / "alternative_topic" / "scideberta" / "final_model",
+        "scideberta": base_dir / "training" / "outputs" / "scideberta_topic" / "final_model",  # BEST: F1=0.478
+        "deberta": base_dir / "training" / "outputs" / "deberta_topic" / "final_model",
+        "scibert": base_dir / "training" / "outputs" / "scibert_topic" / "final_model",
+        "deberta_v3_small": base_dir / "training" / "outputs" / "deberta_v3_small_topic" / "final_model",
+        "modernbert": base_dir / "training" / "outputs" / "modernbert_topic" / "final_model",
+        # Legacy models (pre-Feb 2026, kept for backwards compatibility)
+        "scibert_legacy": base_dir / "scibert" / "scibert_topic" / "final_model",
+        "deberta_legacy": base_dir / "alternative_topic" / "deberta" / "final_model",
+        "scideberta_legacy": base_dir / "alternative_topic" / "scideberta" / "final_model",
     }
     hub_fallback_map = {
-        "scibert": "Sina1138/SciDeberta_Review",
+        "scideberta": "Sina1138/SciDeberta_Review",  # Production HuggingFace model
+        "scibert": "allenai/scibert_scivocab_uncased",  # Needs fine-tuning
+        "deberta": "microsoft/deberta-v3-base",  # Needs fine-tuning
+        "deberta_v3_small": "microsoft/deberta-v3-small",  # Needs fine-tuning
+        "modernbert": "answerdotai/ModernBERT-base",  # Needs fine-tuning
     }
 
     if model_variant not in variant_map:

interface/interactive_processor.py

@@ -64,14 +64,32 @@ class InteractiveReviewProcessor:
         self.rsa_model.eval()
 
         # Load polarity model
-        polarity_model_name = "Sina1138/Scibert_polarity_Review"
+        # Option A (Feb 2026): DeBERTa-v3-base for +5.5% F1 improvement (0.764 vs 0.724 SciBERT)
+        # Try local trained model first, fall back to HuggingFace
+        polarity_model_local = BASE_DIR / "training" / "outputs" / "deberta_polarity" / "final_model"
+        if polarity_model_local.exists() and (polarity_model_local / "config.json").exists():
+            polarity_model_name = str(polarity_model_local)
+            print(f"Loading polarity model from local trained model: {polarity_model_name}")
+        else:
+            # Fallback: will need to upload fine-tuned model or use legacy SciBERT
+            polarity_model_name = "Sina1138/Scibert_polarity_Review"  # Legacy SciBERT
+            print(f"Local model not found, using legacy SciBERT: {polarity_model_name}")
+
         self.polarity_tokenizer = AutoTokenizer.from_pretrained(polarity_model_name)
         self.polarity_model = AutoModelForSequenceClassification.from_pretrained(polarity_model_name)
         self.polarity_model.to(self.device)
         self.polarity_model.eval()
 
         # Load topic model
-        topic_model_name = "Sina1138/SciDeberta_Review"
+        # SciDeBERTa maintains best performance (F1=0.478)
+        topic_model_local = BASE_DIR / "training" / "outputs" / "scideberta_topic" / "final_model"
+        if topic_model_local.exists() and (topic_model_local / "config.json").exists():
+            topic_model_name = str(topic_model_local)
+            print(f"Loading topic model from local trained model: {topic_model_name}")
+        else:
+            topic_model_name = "Sina1138/SciDeberta_Review"  # Production HuggingFace model
+            print(f"Using HuggingFace topic model: {topic_model_name}")
+
         self.topic_tokenizer = AutoTokenizer.from_pretrained(topic_model_name)
         self.topic_model = AutoModelForSequenceClassification.from_pretrained(topic_model_name)
         self.topic_model.to(self.device)

pipeline/config.py

@@ -28,9 +28,26 @@ class Config:
     OPENREVIEW_BASE_URL = 'https://api2.openreview.net'
     VENUE_TEMPLATE = 'ICLR.cc/{year}/Conference'
 
-    # Model paths (HuggingFace)
-    POLARITY_MODEL = "Sina1138/Scibert_polarity_Review"
-    TOPIC_MODEL = "Sina1138/SciDeberta_Review"
+    # Model paths
+    # Option A (Maximize Accuracy): DeBERTa polarity + SciDeBERTa topic - Feb 2026 upgrade
+    # Polarity: DeBERTa-v3-base (F1=0.764, +5.5% vs SciBERT baseline 0.724)
+    # Topic: SciDeBERTa (F1=0.478, maintains lead)
+
+    # Local trained models (preferred for production after validation)
+    POLARITY_MODEL_LOCAL = BASE_DIR / "training" / "outputs" / "deberta_polarity" / "final_model"
+    TOPIC_MODEL_LOCAL = BASE_DIR / "training" / "outputs" / "scideberta_topic" / "final_model"
+
+    # HuggingFace fallbacks (if local models not available)
+    POLARITY_MODEL_HUB = "Sina1138/Scibert_polarity_Review"  # Legacy SciBERT (until fine-tuned DeBERTa is uploaded to Hub)
+    TOPIC_MODEL_HUB = "Sina1138/SciDeberta_Review"  # Current production model
+
+    # Legacy models (SciBERT baseline, kept for reference)
+    POLARITY_MODEL_LEGACY = "Sina1138/Scibert_polarity_Review"  # F1=0.724
+
+    # Use local models if available, otherwise fall back to hub
+    POLARITY_MODEL = str(POLARITY_MODEL_LOCAL) if POLARITY_MODEL_LOCAL.exists() else POLARITY_MODEL_HUB
+    TOPIC_MODEL = str(TOPIC_MODEL_LOCAL) if TOPIC_MODEL_LOCAL.exists() else TOPIC_MODEL_HUB
+
     RSA_MODEL = "sshleifer/distilbart-cnn-12-3"  # For GLIMPSE
 
     @classmethod

pipeline/scored_reviews_builder.py

@@ -6,11 +6,15 @@ from pathlib import Path
 from tqdm import tqdm
 import json
 
-BASE_DIR = Path(__file__).resolve().parent.parent
-sys.path.insert(0, str(BASE_DIR))
+_dir = Path(__file__).resolve().parent
+sys.path.insert(0, str(_dir))
+sys.path.insert(0, str(_dir.parent))
 
+from config import Config
 from dependencies.Glimpse_tokenizer import glimpse_tokenizer
 
+BASE_DIR = Config.BASE_DIR
+
 # def tokenize_sentences(text: str) -> list:
 #     # same tokenization as in the original glimpse code
 #     text = text.replace('-----', '\n')