Update app.py

app.py

@@ -17,6 +17,7 @@ dynamic_temperature_range = 0.15
 rand_range = random.uniform(-1 * dynamic_temperature_range , dynamic_temperature_range )
 temperature = base_temperature  + rand_range
 
+SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"]
 
 #See document for experiment https://www.linkedin.com/pulse/ai-goes-job-interview-g%C3%B6rkem-g%C3%B6knar/
 
@@ -27,7 +28,13 @@ def get_chat_response(name, input_txt = "Hello , what is your name?"):
 
   ##can respond well to history as well but for this quick demo not implemented
   ##see metayazar.com/chatbot for a min 2 history
-  bot_input_ids = tokenizer.encode(personality + tokenizer.eos_token + input_txt + tokenizer.eos_token  , return_tensors='pt')
+  
+  ##this is a multi-speaker model, currently no history, so ending with "<speaker2>" to get response. depends on who starts conversation it can be speaker1
+  ##if there is a history depends on who started conversation it should end with <speaker1>
+  #historical implementation not implemented in this demo
+  
+
+  bot_input_ids =  tokenizer.encode(tokenizer.bos_token + personality + tokenizer.eos_token + input_txt + tokenizer.eos_token + "<speaker2>"   , return_tensors='pt')
 
   #optimum response and speed
   #50 token max length, temperature = 1.3 makes it creative