Summarization fix

scripts/demo_gradio.py

@@ -126,7 +126,8 @@ def predict(text: str, compression: int):
         logger.info("Generating summary with max length %s", max_len)
 
         summary = pipeline.summarize([text], max_length=max_len)[0].strip()
-        emotions = pipeline.predict_emotions([text])[0]
+        # Use a higher threshold to filter out weak/wrong predictions on out-of-domain text
+        emotions = pipeline.predict_emotions([text], threshold=0.6)[0]
         topic = pipeline.predict_topics([text])[0]
 
         fallback_summary = None
@@ -451,10 +452,16 @@ def load_rouge_metrics():
         )
 
     table = pd.DataFrame(rows, columns=columns) if rows else empty
+    
+    # Clean up path for display
+    display_path = str(ROUGE_REPORT_PATH)
+    if "/app/" in display_path:
+        display_path = display_path.replace("/app/", "/LexiMind/")
+        
     metadata = {
         "num_examples": report.get("num_examples"),
         "config": report.get("config"),
-        "report_path": str(ROUGE_REPORT_PATH),
+        "report_path": display_path,
         "last_updated": datetime.fromtimestamp(ROUGE_REPORT_PATH.stat().st_mtime).isoformat(),
     }
     return table, metadata

src/inference/pipeline.py

@@ -75,8 +75,8 @@ class InferencePipeline:
         with torch.inference_mode():
             encoder_mask = src_mask.unsqueeze(1) & src_mask.unsqueeze(2) if src_mask is not None else None
             memory = self.model.encoder(src_ids, mask=encoder_mask)
-            # Relax min_len to avoid forcing repetition if the model wants to stop
-            min_len = 0
+            # Force a minimum length to prevent immediate EOS
+            min_len = 10
             generated = self.model.decoder.greedy_decode(
                 memory=memory,
                 max_len=max_len,
@@ -86,11 +86,18 @@ class InferencePipeline:
                 min_len=min_len,
             )
             
-            # If the first token is EOS, it means empty generation.
-            # Try forcing a different start token if that happens, or just accept it.
-            # For now, we just decode.
+            # Post-process to remove repetition if detected
+            decoded_list = self.tokenizer.decode_batch(generated.tolist())
+            final_summaries = []
+            for summary in decoded_list:
+                # Simple repetition check: if the string starts with a repeated pattern
+                # "TextText" -> "Text" == "Text"
+                if len(summary) > 20 and summary[:4] == summary[4:8]:
+                     final_summaries.append("") # Fallback to empty if garbage
+                else:
+                     final_summaries.append(summary)
             
-        return self.tokenizer.decode_batch(generated.tolist())
+        return final_summaries
 
     def predict_emotions(
         self,