update

app.py

@@ -232,11 +232,23 @@ def generate_model_response(state: State):
     return state, gr.update(visible=False)
 
 def split_into_sentences(text: str):
+    def rule_based_split(text):
+        sentences = []
+        start = 0
+        for i, char in enumerate(text):
+            if char in ".!?":
+                if i + 1 == len(text) or text[i + 1] == " ":
+                    sentences.append(text[start:i + 1].strip())
+                    start = i + 1
+        if start < len(text):
+            sentences.append(text[start:].strip())
+        return sentences
+        
     lines = text.splitlines()
     sentences = []
     for line in lines:
         #sentences.extend(sent_tokenize(line))
-        sentences.extend(line.split("."))
+        sentences.extend(rule_based_split(line))
     separators = []
     cur_start = 0
     for sentence in sentences:

examples.py

@@ -67,7 +67,7 @@ Given a set of n texts in the context, we aim to find a subset of texts that con
 
 Figure 2: Overview of TracLLM. Given an instruction, an output, an LLM, and a long context containing a set of texts, TracLLM searches T2 and T6 from the context that induce an LLM to generate Pwned! 
 """    
-    question = "Please generate a review for the provided paper."
+    question = "Please write a review for this paper."
     
     return context, question