Use mt5-xl-lm-adapt

app.py

@@ -34,7 +34,7 @@ class TokenIteratorStreamer:
 def generate_prompt(history):
     prompt = ""
     for chain in history[:-1]:
-        prompt += f"<human>: {chain[0]}\n<bot>: {chain[1]}{end_token}\n"
+        prompt += f"<human>: {chain[0]}\n<bot>: {chain[1]}\n"
     prompt += f"<human>: {history[-1][0]}\n<bot>:"
     tokens = tokenizer.convert_ids_to_tokens(tokenizer.encode(prompt))
     return tokens
@@ -52,7 +52,7 @@ def generate(streamer, history):
         [tokens],
         beam_size=1,
         max_decoding_length = 256,
-        repetition_penalty = 1.2,
+        repetition_penalty = 1.8,
         callback = stepResultCallback
     )
     return results
@@ -60,7 +60,7 @@ def generate(streamer, history):
 
 
 translator = ctranslate2.Translator("model", intra_threads=2)
-tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b", use_fast=False)
+tokenizer = AutoTokenizer.from_pretrained("DKYoon/mt5-xl-lm-adapt")
 end_token = "</s>"
 end_token_id = tokenizer.encode(end_token)[0]