Merge pull request #28 from Y-IAB/27-instruction

app.py

@@ -3,35 +3,25 @@ It provides a platform for comparing the responses of two LLMs.
 """
 
 import enum
-from random import sample
 from uuid import uuid4
 
 import firebase_admin
 from firebase_admin import firestore
 import gradio as gr
-from litellm import completion
 
 from leaderboard import build_leaderboard
+import response
+from response import get_responses
 
 # TODO(#21): Fix auto-reload issue related to the initialization of Firebase.
 db_app = firebase_admin.initialize_app()
 db = firestore.client()
 
-# TODO(#1): Add more models.
-SUPPORTED_MODELS = [
-    "gpt-4", "gpt-4-0125-preview", "gpt-3.5-turbo", "gemini-pro"
-]
-
 SUPPORTED_TRANSLATION_LANGUAGES = [
     "Korean", "English", "Chinese", "Japanese", "Spanish", "French"
 ]
 
 
-class ResponseType(enum.Enum):
-  SUMMARIZE = "Summarize"
-  TRANSLATE = "Translate"
-
-
 class VoteOptions(enum.Enum):
   MODEL_A = "Model A is better"
   MODEL_B = "Model B is better"
@@ -39,107 +29,40 @@ class VoteOptions(enum.Enum):
 
 
 def vote(vote_button, response_a, response_b, model_a_name, model_b_name,
-         user_prompt, res_type, source_lang, target_lang):
+         user_prompt, instruction, category, source_lang, target_lang):
   doc_id = uuid4().hex
   winner = VoteOptions(vote_button).name.lower()
 
-  if res_type == ResponseType.SUMMARIZE.value:
+  doc = {
+      "id": doc_id,
+      "prompt": user_prompt,
+      "instruction": instruction,
+      "model_a": model_a_name,
+      "model_b": model_b_name,
+      "model_a_response": response_a,
+      "model_b_response": response_b,
+      "winner": winner,
+      "timestamp": firestore.SERVER_TIMESTAMP
+  }
+
+  if category == response.Category.SUMMARIZE.value:
     doc_ref = db.collection("arena-summarizations").document(doc_id)
-    doc_ref.set({
-        "id": doc_id,
-        "prompt": user_prompt,
-        "model_a": model_a_name,
-        "model_b": model_b_name,
-        "model_a_response": response_a,
-        "model_b_response": response_b,
-        "winner": winner,
-        "timestamp": firestore.SERVER_TIMESTAMP
-    })
+    doc_ref.set(doc)
     return
 
-  if res_type == ResponseType.TRANSLATE.value:
+  if category == response.Category.TRANSLATE.value:
     doc_ref = db.collection("arena-translations").document(doc_id)
-    doc_ref.set({
-        "id": doc_id,
-        "prompt": user_prompt,
-        "model_a": model_a_name,
-        "model_b": model_b_name,
-        "model_a_response": response_a,
-        "model_b_response": response_b,
-        "source_language": source_lang.lower(),
-        "target_language": target_lang.lower(),
-        "winner": winner,
-        "timestamp": firestore.SERVER_TIMESTAMP
-    })
-
-
-def response_generator(response: str):
-  for part in response:
-    content = part.choices[0].delta.content
-    if content is None:
-      continue
-
-    # To simulate a stream, we yield each character of the response.
-    for character in content:
-      yield character
-
-
-def get_responses(user_prompt):
-  models = sample(SUPPORTED_MODELS, 2)
-
-  generators = []
-  for model in models:
-    try:
-      # TODO(#1): Allow user to set configuration.
-      response = completion(model=model,
-                            messages=[{
-                                "content": user_prompt,
-                                "role": "user"
-                            }],
-                            stream=True)
-      generators.append(response_generator(response))
-
-    # TODO(#1): Narrow down the exception type.
-    except Exception as e:  # pylint: disable=broad-except
-      print(f"Error in bot_response: {e}")
-      raise e
-
-  responses = ["", ""]
-
-  # It simulates concurrent response generation from two models.
-  while True:
-    stop = True
-
-    for i in range(len(generators)):
-      try:
-        yielded = next(generators[i])
-
-        if yielded is None:
-          continue
-
-        responses[i] += yielded
-        stop = False
-
-        yield responses + models
-
-      except StopIteration:
-        pass
-
-      # TODO(#1): Narrow down the exception type.
-      except Exception as e:  # pylint: disable=broad-except
-        print(f"Error in generator: {e}")
-        raise e
-
-    if stop:
-      break
+    doc["source_lang"] = source_lang.lower()
+    doc["target_lang"] = target_lang.lower()
+    doc_ref.set(doc)
 
 
 with gr.Blocks(title="Arena") as app:
   with gr.Row():
-    response_type_radio = gr.Radio(
-        [response_type.value for response_type in ResponseType],
-        label="Response type",
-        info="Choose the type of response you want from the model.")
+    category_radio = gr.Radio(
+        [category.value for category in response.Category],
+        label="Category",
+        info="The chosen category determines the instruction sent to the LLMs.")
 
     source_language = gr.Dropdown(
         choices=SUPPORTED_TRANSLATION_LANGUAGES,
@@ -154,15 +77,15 @@ with gr.Blocks(title="Arena") as app:
         interactive=True,
         visible=False)
 
-    def update_language_visibility(response_type):
-      visible = response_type == ResponseType.TRANSLATE.value
+    def update_language_visibility(category):
+      visible = category == response.Category.TRANSLATE.value
       return {
           source_language: gr.Dropdown(visible=visible),
           target_language: gr.Dropdown(visible=visible)
       }
 
-    response_type_radio.change(update_language_visibility, response_type_radio,
-                               [source_language, target_language])
+    category_radio.change(update_language_visibility, category_radio,
+                          [source_language, target_language])
 
   model_names = [gr.State(None), gr.State(None)]
   response_boxes = [gr.State(None), gr.State(None)]
@@ -175,7 +98,7 @@ with gr.Blocks(title="Arena") as app:
     response_boxes[1] = gr.Textbox(label="Model B", interactive=False)
 
   # TODO(#5): Display it only after the user submits the prompt.
-  # TODO(#6): Block voting if the response_type is not set.
+  # TODO(#6): Block voting if the category is not set.
   # TODO(#6): Block voting if the user already voted.
   with gr.Row():
     option_a = gr.Button(VoteOptions.MODEL_A.value)
@@ -188,10 +111,15 @@ with gr.Blocks(title="Arena") as app:
       model_names[0] = gr.Textbox(label="Model A", interactive=False)
       model_names[1] = gr.Textbox(label="Model B", interactive=False)
 
-  submit.click(get_responses, prompt, response_boxes + model_names)
+  instruction_state = gr.State("")
+
+  submit.click(get_responses,
+               [prompt, category_radio, source_language, target_language],
+               response_boxes + model_names + [instruction_state])
 
   common_inputs = response_boxes + model_names + [
-      prompt, response_type_radio, source_language, target_language
+      prompt, instruction_state, category_radio, source_language,
+      target_language
   ]
   option_a.click(vote, [option_a] + common_inputs)
   option_b.click(vote, [option_b] + common_inputs)

response.py

@@ -0,0 +1,100 @@
+"""
+This module contains functions for generating responses using LLMs.
+"""
+
+import enum
+from random import sample
+
+import gradio as gr
+from litellm import completion
+
+# TODO(#1): Add more models.
+SUPPORTED_MODELS = [
+    "gpt-4", "gpt-4-0125-preview", "gpt-3.5-turbo", "gemini-pro"
+]
+
+
+class Category(enum.Enum):
+  SUMMARIZE = "Summarize"
+  TRANSLATE = "Translate"
+
+
+# TODO(#31): Let the model builders set the instruction.
+def get_instruction(category, source_lang, target_lang):
+  if category == Category.SUMMARIZE.value:
+    return "Summarize the following text in its original language."
+  if category == Category.TRANSLATE.value:
+    return f"Translate the following text from {source_lang} to {target_lang}."
+
+
+def response_generator(response: str):
+  for part in response:
+    content = part.choices[0].delta.content
+    if content is None:
+      continue
+
+    # To simulate a stream, we yield each character of the response.
+    for character in content:
+      yield character
+
+
+# TODO(#29): Return results simultaneously to prevent bias from generation speed.
+def get_responses(user_prompt, category, source_lang, target_lang):
+  if not category:
+    raise gr.Error("Please select a category.")
+
+  if category == Category.TRANSLATE.value and (not source_lang or
+                                               not target_lang):
+    raise gr.Error("Please select source and target languages.")
+
+  models = sample(SUPPORTED_MODELS, 2)
+  instruction = get_instruction(category, source_lang, target_lang)
+
+  generators = []
+  for model in models:
+    try:
+      # TODO(#1): Allow user to set configuration.
+      response = completion(model=model,
+                            messages=[{
+                                "content": instruction,
+                                "role": "system"
+                            }, {
+                                "content": user_prompt,
+                                "role": "user"
+                            }],
+                            stream=True)
+      generators.append(response_generator(response))
+
+    # TODO(#1): Narrow down the exception type.
+    except Exception as e:  # pylint: disable=broad-except
+      print(f"Error in bot_response: {e}")
+      raise e
+
+  responses = ["", ""]
+
+  # It simulates concurrent response generation from two models.
+  while True:
+    stop = True
+
+    for i in range(len(generators)):
+      try:
+        yielded = next(generators[i])
+
+        if yielded is None:
+          continue
+
+        responses[i] += yielded
+        stop = False
+
+        yield responses + models + [instruction]
+
+      except StopIteration:
+        pass
+
+      # TODO(#1): Narrow down the exception type.
+      except Exception as e:  # pylint: disable=broad-except
+        print(f"Error in generator: {e}")
+        raise e
+
+    if stop:
+      break