Update chat interface with feedback Dani

README.md

@@ -1,5 +1,5 @@
 ---
-title: LLM Human Feedback Collector Chat Interface (DPO)
+title: LLM Human Feedback Collector | Meta-Llama-3.1-8B-Instruct | (DPO)
 emoji: 🦾💪🏽
 colorFrom: pink
 colorTo: blue
@@ -12,4 +12,4 @@ suggested_hardware: t4-small
 short_description: LLM, chatbot
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Check out the configuration reference at <https://huggingface.co/docs/hub/spaces-config-reference>

app.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python
 import os
-from threading import Thread
 from typing import Iterator
 
 import gradio as gr
@@ -15,10 +14,14 @@ DEFAULT_MAX_NEW_TOKENS = 1024
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "8192"))
 
 if torch.cuda.is_available():
-    model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
-    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
-    tokenizer = AutoTokenizer.from_pretrained(model_id)
-style = "<style>.user-message,.system-message{display:flex;margin:10px}.user-message .message-content{background-color:#c2e3f7;color:#000}.system-message .message-content{background-color:#f5f5f5;color:#000}.message-content{padding:10px;border-radius:10px;max-width:70%;word-wrap:break-word}.container{display:flex;justify-content:space-between}.column{width:48%}</style>"
+    # model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
+    # model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
+    # tokenizer = AutoTokenizer.from_pretrained(model_id)
+    pass
+
+style = None
+
+AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 
 
 @spaces.GPU
@@ -31,36 +34,39 @@ def generate(
     top_k: int = 40,
     repetition_penalty: float = 1.2,
 ) -> Iterator[str]:
-    conversation = []
-    for user, assistant in chat_history:
-        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
-    conversation.append({"role": "user", "content": message})
-
-    input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
-    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
-        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
-        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
-    input_ids = input_ids.to(model.device)
-
-    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
-    generate_kwargs = dict(
-        {"input_ids": input_ids},
-        streamer=streamer,
-        max_new_tokens=max_new_tokens,
-        do_sample=True,
-        top_p=top_p,
-        top_k=top_k,
-        temperature=temperature,
-        num_beams=1,
-        repetition_penalty=repetition_penalty,
-    )
-    t = Thread(target=model.generate, kwargs=generate_kwargs)
-    t.start()
-
-    outputs = []
-    for text in streamer:
-        outputs.append(text)
-        yield "".join(outputs)
+    # conversation = []
+    # for user, assistant in chat_history:
+    #     conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
+    # conversation.append({"role": "user", "content": message})
+
+    # input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
+    # if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
+    #     input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
+    #     gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
+    # input_ids = input_ids.to(model.device)
+
+    # streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
+    # generate_kwargs = dict(
+    #     {"input_ids": input_ids},
+    #     streamer=streamer,
+    #     max_new_tokens=max_new_tokens,
+    #     do_sample=True,
+    #     top_p=top_p,
+    #     top_k=top_k,
+    #     temperature=temperature,
+    #     num_beams=1,
+    #     repetition_penalty=repetition_penalty,
+    # )
+    # t = Thread(target=model.generate, kwargs=generate_kwargs)
+    # t.start()
+
+    # outputs = []
+    # for text in streamer:
+    #     outputs.append(text)
+    #     yield "".join(outputs)
+
+    for char in "help":
+        yield "help"
 
 
 chat_interface = ChatInterface(
@@ -69,10 +75,7 @@ chat_interface = ChatInterface(
     min_turns=1,
     max_turns=10,
     repo_id="llm-human-feedback-collector-chat-interface-dpo",
-    chatbot=gr.Chatbot(
-        height=450, label="Meta-Llama-3.1-8B-Instruct", show_share_button=True
-    ),
-    css=style,
+    chatbot=gr.Chatbot(height=450, label="Meta-Llama-3.1-8B-Instruct", show_share_button=True),
     cache_examples=False,
     additional_inputs=[
         gr.Slider(

chat_interface_preference.py

@@ -144,15 +144,15 @@ class ChatInterface(Blocks):
         submit_btn_bad = None
         stop_btn = "Stop"
         undo_btn = "↩️ Undo"
-        clear_btn = "🗑️ Log and clear"
+        clear_btn = "🗑️ Clear"
         if "kto" in prefence_techniques:
-            submit_btn_good = "Log response 👍"
-            submit_btn_bad = "Log response 👎"
+            submit_btn_good = "The response 👍"
+            submit_btn_bad = "The response 👎"
         if any([technique for technique in ["dpo", "simpo", "rlhf", "orpo"] if technique in self.prefence_techniques]):
-            submit_btn_two = "Generate 2"
-            submit_btn_a = "Log preference 🅰️"
-            submit_btn_b = "Log preference 🅱️"
-            submit_btn_ab = "Continue random 🅰️=🅱️"
+            submit_btn_two = None
+            submit_btn_a = "A is better than B"
+            submit_btn_b = "B is better than A"
+            submit_btn_ab = "A and B are similar"
         super().__init__(
             analytics_enabled=analytics_enabled,
             mode="chat_interface",
@@ -387,13 +387,13 @@ class ChatInterface(Blocks):
 
     def _setup_events(self) -> None:
         submit_fn_one = self._stream_fn if self.is_generator else self._submit_fn
-        submit_fn_one_partial = functools.partial(submit_fn_one, n_generations=1)
+        submit_fn_one_partial = functools.partial(submit_fn_one, n_generations=2)
         submit_triggers_one = (
             [self.textbox.submit, self.submit_btn_one.click] if self.submit_btn_one else [self.textbox.submit]
         )
         submit_tuples = [(submit_fn_one_partial, submit_triggers_one)]
         if self.submit_btn_two:
-            submit_fn_two = functools.partial(submit_fn_one, n_generations=2)
+            submit_fn_two = functools.partial(submit_fn_one, n_generations=1)
             submit_triggers_two = [self.submit_btn_two.click]
             submit_tuples.append((submit_fn_two, submit_triggers_two))
         for _fn, _triggers in submit_tuples:
@@ -608,7 +608,7 @@ class ChatInterface(Blocks):
             if turn[-1]:
                 conversation += self._get_chat_message(turn[-1], role="user", turn=(idx + 1))
 
-        return self.css + "<body>" + conversation + "</body>"
+        return "<body>" + self.css + conversation + "</body>"
 
     def _get_conversation_in_openai_format(self, history):
         conversation = []
@@ -622,26 +622,22 @@ class ChatInterface(Blocks):
 
     @staticmethod
     def _get_chat_message(message, role, turn):
-        if role == "user":
-            justify = "right"
-        else:
-            justify = "left"
+        # return f"<p><div class='message-identifier'>{message}</div></p>"
         return (
-            f'<div class="{role}-message" style="justify-content: {justify};">'
-            + '<div class="message-content">'
-            + f"<strong>Turn {turn} - {role.capitalize()}:</strong><br>"
+            '<div class="message-content">'
+            + f"<strong>Option {turn} - </strong><br>"
             + f"<em>Length: {len(message)} characters</em><br><br>"
             + f'<div class="message-identifier">{message}</div>'
-            + "</div></div>"
+            + "</div>"
         )
 
     def _get_chat_message_comparison(self, content_a, content_b):
         return (
-            '<div class="container">'
-            + '<div class="column">'
+            '<div class="container" style="display: flex; width: 100%;">'
+            + '<div class="column" style="flex: 1; padding: 10px;">'
             + self._get_chat_message(message=content_a, role="system", turn="A")
             + "</div>"
-            + '<div class="column">'
+            + '<div class="column" style="flex: 1; padding: 10px;">'
             + self._get_chat_message(message=content_b, role="system", turn="B")
             + "</div>"
             + "</div>"

requirements.txt

@@ -1,8 +1,10 @@
-accelerate==0.31.0
-bitsandbytes==0.42
-gradio==4.36.1
+gradio==4.39
 scipy==1.13.0
-sentencepiece==0.2.0
 spaces==0.28.3
 torch==2.0.1
-transformers==4.41.2
+accelerate
+bitsandbytes
+torch
+transformers==4.43.1
+einops
+sentencepiece