feat: add NTU LLaMA model

app.py

@@ -15,7 +15,6 @@ from utils.utils import (
 )
 from utils.completion_reward import CompletionReward
 from utils.completion_reward_utils import (
-    get_llm_response,
     set_player_name,
     set_player_selected_character,
     create_certificate,
@@ -258,6 +257,10 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
             mtk_description = gr.Markdown(
                 "# 蔚藍", elem_id="mtk_description", visible=False
             )
+            ntu_description = gr.Markdown(
+                "# 紅寶石", elem_id="ntu_description", visible=False
+            )
+
 
         with gr.Row():
             openai_img = gr.Image(
@@ -288,6 +291,13 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
                 interactive=False,
                 show_download_button=False,
             )
+            ntu_img = gr.Image(
+                "medias/ntu.png",
+                visible=False,
+                elem_id="ntu_img",
+                interactive=False,
+                show_download_button=False,
+            )
 
         with gr.Row():
             start_generate_story = gr.Button(
@@ -306,6 +316,7 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
             bot2 = gr.Chatbot(visible=False, elem_id="bot2")
             bot3 = gr.Chatbot(visible=False, elem_id="bot3")
             bot4 = gr.Chatbot(visible=False, elem_id="bot4")
+            bot5 = gr.Chatbot(visible=False, elem_id="bot5")
 
         with gr.Row():
             select_story = gr.Radio(
@@ -460,31 +471,26 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
         ],
         queue=False,
     ).then(
-        lambda: create_visibility_updates(True, 11),
+        lambda: create_visibility_updates(True, 13),
         None,
         [
             openai_img,
             aws_img,
             google_img,
             mtk_img,
+            ntu_img,
             story_title,
             story_description,
             openai_description,
             aws_description,
             google_description,
             mtk_description,
+            ntu_description,
             start_generate_story,
         ],
         queue=False,
     )
 
-    get_llm_response_args = dict(
-        fn=get_llm_response,
-        inputs=[completion_reward, player_logs],
-        outputs=[bot1, bot2, bot3, bot4],
-        queue=False,
-    )
-
     get_first_llm_response_args = dict(
         fn=get_llm_response_once,
         inputs=[completion_reward, player_logs],
@@ -513,12 +519,19 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
         queue=False,
     )
 
+    get_fifth_llm_response_args = dict(
+        fn=get_llm_response_once,
+        inputs=[completion_reward, player_logs],
+        outputs=bot5,
+        queue=False,
+    )
+
     start_generate_story.click(
         lambda: gr.update(visible=False), None, start_generate_story, queue=False
     ).then(
-        lambda: create_visibility_updates(True, 5),
+        lambda: create_visibility_updates(True, 6),
         None,
-        [bot1, bot2, bot3, bot4, weaving],
+        [bot1, bot2, bot3, bot4, bot5, weaving],
         queue=False,
     ).then(
         **get_first_llm_response_args
@@ -528,6 +541,8 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
         **get_third_llm_response_args
     ).then(
         **get_fourth_llm_response_args
+    ).then(
+        **get_fifth_llm_response_args
     ).then(
         lambda: gr.update(visible=True), None, [select_story], queue=False
     ).then(
@@ -580,23 +595,26 @@ with gr.Blocks(theme=seafoam, css=get_content("css/style.css")) as demo:
     )
 
     start_generate_certificate.click(
-        lambda: create_visibility_updates(False, 18),
+        lambda: create_visibility_updates(False, 21),
         None,
         [
             openai_img,
             aws_img,
             google_img,
             mtk_img,
+            ntu_img,
             story_title,
             story_description,
             openai_description,
             aws_description,
             google_description,
             mtk_description,
+            ntu_description,
             bot1,
             bot2,
             bot3,
             bot4,
+            bot5,
             select_story,
             processing,
             cancel_story,

css/style.css

@@ -374,6 +374,11 @@ input[type="range"]::-ms-track {
     border: None !important;
 }
 
+#bot4 .message {
+    background: rgba(227, 66, 104, 0.4) !important;
+    border: None !important;
+}
+
 #processing {
     margin: 20vh 10vw;
     height: 30vh;
@@ -434,4 +439,8 @@ input[type="range"]::-ms-track {
 
 [data-testid="蔚藍-radio-label"] {
     background: rgba(66, 130, 227, 0.4) !important;
+}
+
+[data-testid="紅寶石-radio-label"] {
+    background: rgba(227, 66, 104, 0.4) !important;
 }

utils/completion_reward.py

@@ -48,17 +48,20 @@ class CompletionReward:
         self.paragraph_aws = None
         self.paragraph_google = None
         self.paragraph_mtk = None
+        self.paragraph_ntu = None
         self.player_certificate_url = None
         self.openai_agent = OpenAIAgent()
         self.aws_agent = AWSAgent()
         self.google_agent = GoogleAgent()
         self.mtk_agent = MTKAgent()
+        self.ntu_agent = NTUAgent()
         self.agents_responses = {}
         self.agent_list = [
             self.openai_agent,
             self.aws_agent,
             self.google_agent,
             self.mtk_agent,
+            self.ntu_agent,
         ]
         self.shuffled_response_order = {}
         self.pop_response_order = []
@@ -67,38 +70,9 @@ class CompletionReward:
             "aws": self.paragraph_aws,
             "google": self.paragraph_google,
             "mtk": self.paragraph_mtk,
+            "ntu": self.paragraph_ntu,
         }
 
-    def get_llm_response(self, player_logs):
-        openai_story = self.openai_agent.get_story(player_logs)
-        aws_story = self.aws_agent.get_story(player_logs)
-        google_story = self.google_agent.get_story(player_logs)
-        mtk_story = self.mtk_agent.get_story(player_logs)
-        agents_responses = {
-            "openai": openai_story,
-            "aws": aws_story,
-            "google": google_story,
-            "mtk": mtk_story,
-        }
-        self.paragraph_openai = agents_responses["openai"]
-        self.paragraph_aws = agents_responses["aws"]
-        self.paragraph_google = agents_responses["google"]
-        self.paragraph_mtk = agents_responses["mtk"]
-        response_items = list(agents_responses.items())
-        random.shuffle(response_items)
-
-        self.shuffled_response_order = {
-            str(index): agent for index, (agent, _) in enumerate(response_items)
-        }
-
-        shuffled_responses = tuple(response for _, response in response_items)
-        return (
-            [(None, shuffled_responses[0])],
-            [(None, shuffled_responses[1])],
-            [(None, shuffled_responses[2])],
-            [(None, shuffled_responses[3])],
-        )
-
     def get_llm_response_once(self, player_logs):
         if self.agent_list:
             # Randomly select and remove an agent from the list
@@ -110,7 +84,7 @@ class CompletionReward:
         self.agents_responses[agent.name] = story
         self.pop_response_order.append(agent.name)
 
-        if len(self.pop_response_order) == 4:
+        if len(self.pop_response_order) == 5:
             self.shuffled_response_order = {
                 str(index): agent for index, agent in enumerate(self.pop_response_order)
             }
@@ -118,6 +92,7 @@ class CompletionReward:
             self.paragraph_aws = self.agents_responses["aws"]
             self.paragraph_google = self.agents_responses["google"]
             self.paragraph_mtk = self.agents_responses["mtk"]
+            self.paragraph_ntu = self.agents_responses["ntu"]
 
         return [(None, story)]
 
@@ -137,6 +112,7 @@ class CompletionReward:
             "索拉拉": "1",
             "薇丹特": "2",
             "蔚藍": "3",
+            "紅寶石": "4",
         }
         self.player_selected_character = player_selected_character
         self.player_selected_model = self.shuffled_response_order[
@@ -290,6 +266,8 @@ class OpenAIAgent:
                 logging.error(f"OpenAI Attempt {retry_attempts}: {e}")
                 time.sleep(1 * retry_attempts)
 
+        return '星際夥伴短時間內寫了太多故事，需要休息一下，請稍後再試，或是選擇其他星際夥伴的故事。'
+
     def get_background(self):
         client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
         image_url = None
@@ -355,6 +333,8 @@ class AWSAgent:
                 logging.error(f"AWS Attempt {retry_attempts}: {e}")
                 time.sleep(1 * retry_attempts)
 
+        return '星際夥伴短時間內寫了太多故事，需要休息一下，請稍後再試，或是選擇其他星際夥伴的故事。'
+
 
 class GoogleAgent:
     from google.cloud import aiplatform
@@ -412,6 +392,8 @@ class GoogleAgent:
                 logging.error(f"Google Attempt {retry_attempts}: {e}")
                 time.sleep(1 * retry_attempts)
 
+        return '星際夥伴短時間內寫了太多故事，需要休息一下，請稍後再試，或是選擇其他星際夥伴的故事。'
+
 
 class MTKAgent:
     def __init__(self):
@@ -485,7 +467,76 @@ class MTKAgent:
                 retry_attempts += 1
                 logging.error(f"MTK Attempt {retry_attempts}: {e}")
                 time.sleep(1 * retry_attempts)
+    
+        return '星際夥伴短時間內寫了太多故事，需要休息一下，請稍後再試，或是選擇其他星際夥伴的故事。'
+
+class NTUAgent:
+    def __init__(self):
+        self.name = "ntu"
+
+    def get_story(self, user_log):
+        system_prompt = """
+            我正在舉辦一個學習型的活動，我為學生設計了一個獨特的故事機制，每天每個學生都會收到屬於自己獨特的冒險紀錄，現在我需要你協助我將這些冒險紀錄，製作成一段冒險故事，請
+            - 以「你」稱呼學生
+            - 可以裁減內容以將內容限制在 1024 個 token 內
+            - 試著合併故事記錄成一段連貫、有吸引力的故事
+            - 請使用 zh_TW
+            - 請直接回覆故事內容，不需要回覆任何訊息
+        """
+
+        user_log = f"""
+            ```{user_log}
+            ```
+        """
+
+        url = 'http://api.twllm.com:20002/v1/chat/completions'
+
+        data = {
+            "model": "yentinglin/Taiwan-LLM-13B-v2.0-chat",
+            "messages": f"{system_prompt}, 以下是我的冒險故事 ```{user_log}```",
+            "temperature": 0.7,
+            "top_p": 1,
+            "n": 1,
+            "max_tokens": 2048,
+            "stop": ["string"],
+            "stream": False,
+            "presence_penalty": 0,
+            "frequency_penalty": 0,
+            "user": "string",
+            "best_of": 1,
+            "top_k": -1,
+            "ignore_eos": False,
+            "use_beam_search": False,
+            "stop_token_ids": [0],
+            "skip_special_tokens": True,
+            "spaces_between_special_tokens": True,
+            "add_generation_prompt": True,
+            "echo": False,
+            "repetition_penalty": 1,
+            "min_p": 0
+        }
+
+        headers = {
+            'accept': 'application/json',
+            'Content-Type': 'application/json'
+        }
+
+        retry_attempts = 0
+        while retry_attempts < 5:
+            try:
+                response = requests.post(url, headers=headers, data=json.dumps(data)).json()
+                response_text = response["choices"][0]["message"]["content"]
+
+                chinese_converter = OpenCC("s2tw")
+
+                return chinese_converter.convert(response_text)
+
+            except Exception as e:
+                retry_attempts += 1
+                logging.error(f"NTU Attempt {retry_attempts}: {e}")
+                time.sleep(1 * retry_attempts)
 
+        return '星際夥伴短時間內寫了太多故事，需要休息一下，請稍後再試，或是選擇其他星際夥伴的故事。'
 
 class ImageProcessor:
     @staticmethod

utils/completion_reward_utils.py

@@ -2,11 +2,6 @@ import json
 
 import gradio as gr
 
-
-def get_llm_response(completion_reward, *args):
-    return completion_reward.get_llm_response(*args)
-
-
 def get_llm_response_once(completion_reward, *args):
     return completion_reward.get_llm_response_once(*args)