Update app.py

app.py

@@ -57,15 +57,42 @@ class PaperScorer:
         self.prompt_template = (
             "Given a certain paper, Title: {title}\nAbstract: {abstract}\nEvaluate the quality of this paper:"
         )
-
+    def _rescale_score(self, s: float) -> float:
+        """
+        Piecewise linear rescaling:
+          0.0 ~ 0.2  ->  0.0 ~ 0.5
+          0.2 ~ 0.6  ->  0.5 ~ 1.0
+          >0.6       ->  1.0
+        """
+        s = max(0.0, min(1.0, s))  # clamp input to [0,1]
+    
+        if s <= 0.2:
+            # scale [0,0.2] to [0,0.5]
+            y = (s / 0.2) * 0.5
+        elif s <= 0.6:
+            # scale (0.2,0.6] to (0.5,1.0]
+            y = 0.5 + ((s - 0.2) / 0.4) * 0.5
+        else:
+            # everything above 0.6 maps to 1.0
+            y = 1.0
+    
+        return round(y, 4)
     def score(self, title: str, abstract: str) -> float:
         prompt = self.prompt_template.format(title=title.strip(), abstract=abstract.strip())
-        inputs = self.tokenizer(prompt, return_tensors='pt', padding=True, truncation=True,
-                                max_length=self.max_length).to(self.device)
+        inputs = self.tokenizer(
+            prompt,
+            return_tensors='pt',
+            padding=True,
+            truncation=True,
+            max_length=self.max_length
+        ).to(self.device)
+
         with torch.no_grad():
-            logits = self.model(**inputs).logits + 2.0 # # NAIPv2 only learned relative magnitudes. To make the values more comparable, we added a compensation value derived from statistics to the output of v2.
-            score = torch.sigmoid(logits).view(-1).item()
-        return round(score, 4)
+            logits = self.model(**inputs).logits
+            raw_score = torch.sigmoid(logits).view(-1).item()
+
+        final_score = self._rescale_score(raw_score) # 
+        return round(final_score, 4)
 
 @spaces.GPU(duration=60, enable_queue=True)
 def predict_v2(title, abstract):
@@ -187,8 +214,15 @@ with gr.Blocks() as iface:
             - It is intended as a tool **for research and educational purposes only**.
             - Please refrain from deliberately embellishing the title and abstract to boost scores, and **avoid making false claims**.
             - The **predicted value** is a probability generated by the model and **does NOT reflect the paper's true quality or novelty**.  
-            - Empirically, it is considered that a predicted influence score greater than 0.60 in v1 indicates a potentially impactful paper, while a score greater than 0.5 in v2 indicates a high-quality paper.
             - The **author takes NO responsibility** for the prediction results.
+            - This demo explores LLMs for paper quality assessment, but is **not optimized against prompt injection attacks**.
+            - For **NAIPv1**, a predicted influence score greater than 0.60 is considered to indicate a potentially impactful paper.  
+            - For **NAIPv2**, a predicted influence score greater than 0.50 is considered to indicate a potentially high-quality paper.  
+              *Note*: Since NAIPv2 was trained without absolute scale supervision, the 0–1 scores are rescaled.  
+              - 0.50 indicates above the statistical average quality of AI papers.  
+              - >0.65 is considered to reach the NeurIPS poster level.  
+              - ~0.75 corresponds to the average NeurIPS spotlight level.  
+              - ~0.85 corresponds to the average NeurIPS oral level.  
             """)
 
     title_input.change(