Update app.py

app.py

@@ -6,64 +6,55 @@ from threading import Thread
 import re
 import uuid
 
-# ------------- MODEL LOADING -------------
-model_name = "FractalAIResearch/Fathom-R1-14B"  # CHANGE this if your org name or model is different
+# ------------ MODEL SETUP ------------
+model_name = "FractalAIResearch/Fathom-R1-14B"  # adjust if using a local path
+device = "cuda" if torch.cuda.is_available() else "cpu"
 
-device = "cuda:0" if torch.cuda.is_available() else "cpu"
-
-model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype="auto")
+model = AutoModelForCausalLM.from_pretrained(
+    model_name, torch_dtype=torch.float16, device_map="auto"
+)
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 
-# -------- MATH FORMATTING UTILITIES --------
+# ------------ MATH FORMATTING ------------
 def format_math(text):
     text = re.sub(r"\[(.*?)\]", r"$$\1$$", text, flags=re.DOTALL)
     text = text.replace(r"\(", "$").replace(r"\)", "$")
     return text
 
+# ------------ SESSION & HISTORY MGMT ------------
 def generate_conversation_id():
     return str(uuid.uuid4())[:8]
 
-# --------- PROMPT CONSTRUCTION ---------
-def construct_prompt(history_state, user_message, system_message):
-    # Adjust custom tags as per your Fathom R1 model prompt format. This is a typical chat prompt pattern.
-    prompt = f"[SYSTEM]\n{system_message.strip()}\n"
-    for m in history_state:
-        if m["role"] == "user":
-            prompt += f"[USER]\n{m['content']}\n"
-        elif m["role"] == "assistant":
-            prompt += f"[ASSISTANT]\n{m['content']}\n"
-    prompt += f"[USER]\n{user_message.strip()}\n[ASSISTANT]\n"
-    return prompt
-
-# --------- SYSTEM PROMPT SETUP ---------
-system_message = (
-    "Your role as an assistant involves thoroughly exploring questions through a systematic thinking process before providing the final precise and accurate solutions. "
-    "Please structure your response into two main sections: Thought and Solution using the specified format: <think> ... </think> ... {Solution section}. "
-    "In the Thought section, detail your reasoning in steps. Then, systematically present the final solution that is logical and concise."
-)
-
-# ----------- GENERATION FUNCTION -----------
+# ------------ GENERATION FUNCTION ------------
 @spaces.GPU(duration=60)
-def generate_response(
-    user_message, max_tokens, temperature, top_p, history_state
-):
+def generate_response(user_message, max_tokens, temperature, top_p, history_state):
     if not user_message.strip():
         return history_state, history_state
 
-    prompt = construct_prompt(history_state, user_message, system_message)
-    inputs = tokenizer(prompt, return_tensors="pt").to(device)
-    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
-
-    generation_kwargs = dict(
-        input_ids=inputs["input_ids"],
-        attention_mask=inputs["attention_mask"],
-        max_new_tokens=int(max_tokens),
-        do_sample=True,
-        temperature=float(temperature),
-        top_p=float(top_p),
-        streamer=streamer,
+    system_prompt = (
+        "You are an advanced math and science assistant developed by Fractal AI Research. "
+        "You solve problems step-by-step with detailed reasoning. Think deeply and clearly."
     )
 
+    prompt = f"System: {system_prompt}\n"
+    for m in history_state:
+        role = "User" if m["role"] == "user" else "Assistant"
+        prompt += f"{role}: {m['content'].strip()}\n"
+    prompt += f"User: {user_message.strip()}\nAssistant:"
+
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True)
+
+    generation_kwargs = {
+        "input_ids": inputs["input_ids"],
+        "attention_mask": inputs["attention_mask"],
+        "max_new_tokens": int(max_tokens),
+        "temperature": temperature,
+        "top_p": top_p,
+        "do_sample": True,
+        "streamer": streamer,
+    }
+
     thread = Thread(target=model.generate, kwargs=generation_kwargs)
     thread.start()
 
@@ -72,23 +63,23 @@ def generate_response(
         {"role": "user", "content": user_message},
         {"role": "assistant", "content": ""}
     ]
-    for new_token in streamer:
-        assistant_response += new_token
-        # Optionally, you can call format_math() on assistant_response
+
+    for token in streamer:
+        assistant_response += token
         new_history[-1]["content"] = assistant_response.strip()
         yield new_history, new_history
 
     yield new_history, new_history
 
-# ------------ EXAMPLE MESSAGES --------------
+# ------------ EXAMPLES ------------
 example_messages = {
-    "IIT-JEE 2024 Mathematics": "A student appears for a quiz consisting of only true-false type questions and answers all the questions. The student knows the answers of some questions and guesses the answers for the remaining questions. Whenever the student knows the answer of a question, he gives the correct answer. Assume that probability of the student giving the correct answer for a question, given that he has guessed it, is $\\frac{1}{2}$. Also assume that the probability of the answer for a question being guessed, given that the student's answer is correct, is $\\frac{1}{6}$. Then the probability that the student knows the answer of a randomly chosen question is?",
-    "IIT-JEE 2025         Physics": "A person sitting inside an elevator performs a weighing experiment with an object of mass 50 kg. Suppose that the variation of the height 𝑦 (in m) of the elevator, from the ground, with time 𝑡 (in s) is given by 𝑦 = 8 [1 + sin ( 2𝜋𝑡/𝑇 )], where 𝑇 = 40𝜋 s. Taking acceleration due to gravity, 𝑔 = 10 m/s^2 , the maximum variation of the object’s weight (in N) as observed in the experiment is ?",
-    "Goldman Sachs Interview Puzzle": "Four friends need to cross a dangerous bridge at night. Unfortunately, they have only one torch and the bridge is too dangerous to cross without one. The bridge is only strong enough to support two people at a time. Not all people take the same time to cross the bridge. Times for each person: 1 min, 2 mins, 7 mins and 10 mins. What is the shortest time needed for all four of them to cross the bridge?",
-    "IIT-JEE 2025 Mathematics": "Let 𝑆 be the set of all seven-digit numbers that can be formed using the digits 0, 1 and 2. For example, 2210222 is in 𝑆, but 0210222 is NOT in 𝑆.Then the number of elements 𝑥 in 𝑆 such that at least one of the digits 0 and 1 appears exactly twice in 𝑥, is ?"
+    "IIT-JEE 2024 Mathematics": "A student appears for a quiz consisting of only true-false type questions...",
+    "IIT-JEE 2025         Physics": "A person sitting inside an elevator performs a weighing experiment...",
+    "Goldman Sachs Interview Puzzle": "Four friends need to cross a dangerous bridge at night...",
+    "IIT-JEE 2025 Mathematics": "Let 𝑆 be the set of all seven-digit numbers that can be formed using the digits 0, 1 and 2..."
 }
 
-# --------------- UI (GRADIO) ----------------
+# ------------ UI LAYOUT ------------
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
     conversations_state = gr.State({})
     current_convo_id = gr.State(generate_conversation_id())
@@ -98,7 +89,7 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         """
         <div style="display: flex; align-items: center; gap: 16px; margin-bottom: 1em;">
             <div style="background-color: black; padding: 6px; border-radius: 8px;">
-                <img src="https://framerusercontent.com/images/j0KjQQyrUfkFw4NwSaxQOLAoBU.png" style="height: 48px;">
+                <img src="https://framerusercontent.com/images/j0KjQQyrUfkFw4NwSaxQOLAoBU.png" alt="Fractal AI Logo" style="height: 48px;">
             </div>
             <h1 style="margin: 0;">Fathom R1 14B Chatbot</h1>
         </div>
@@ -112,34 +103,84 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
 
     with gr.Row():
         with gr.Column(scale=1):
-            gr.Markdown(
-                """
-                Welcome to the Fathom R1 14B Chatbot, developed by Fractal AI Research! 
-                
-                Our model excels at reasoning tasks in mathematics and science. Given that our model has been optimised for tasks requiring critical thinking, it might overthink for simple chat queries.
-
-                To check out our GitHub repository, click [here](https://github.com/FractalAIResearchLabs/Fathom-R1) 
-                
-                For training recipe details on how this model was built, please check [here](https://huggingface.co/FractalAIResearch/Fathom-R1-14B)  
-                
-                Try the example problems below from various popular entrance examinations and interviews or type in your own problems to see how our model breaks down and solves complex reasoning problems.
-
-                NOTE: Once you close this demo window, all currently saved conversations will be lost.
-                """
-            )
+            gr.Markdown("""
+            Welcome to the Fathom R1 14B Chatbot, developed by Fractal AI Research!
+
+            Our model excels at reasoning tasks in mathematics and science. Given that our model has been optimised for tasks requiring critical thinking, it might overthink for simple chat queries.
+
+            To check out our GitHub repository, click [here](https://github.com/FractalAIResearchLabs/Fathom-R1)
+
+            For training recipe details on how this model was built, please check [here](https://huggingface.co/FractalAIResearch/Fathom-R1-14B)
+
+            Try the example problems below from various popular entrance examinations and interviews or type in your own problems to see how our model breaks down and solves complex reasoning problems.
+            """)
+
             gr.Markdown("### Settings")
             max_tokens_slider = gr.Slider(minimum=6144, maximum=32768, step=1024, value=16384, label="Max Tokens")
             with gr.Accordion("Advanced Settings", open=True):
                 temperature_slider = gr.Slider(minimum=0.1, maximum=2.0, value=0.6, label="Temperature")
                 top_p_slider = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, label="Top-p")
-            gr.Markdown(
-                """
-                We sincerely acknowledge [VIDraft](https://huggingface.co/VIDraft) for their Phi 4 Reasoning Plus [space](https://huggingface.co/spaces/VIDraft/phi-4-reasoning-plus), which served as the starting point for this demo.
-                """
-            )
+
+            gr.Markdown("We sincerely acknowledge [VIDraft](https://huggingface.co/VIDraft) for their Phi 4 Reasoning Plus [space](https://huggingface.co/spaces/VIDraft/phi-4-reasoning-plus), which served as the starting point for this demo.")
+
         with gr.Column(scale=4):
             chatbot = gr.Chatbot(label="Chat", type="messages", height=520)
             with gr.Row():
                 user_input = gr.Textbox(label="User Input", placeholder="Type your question here...", lines=3, scale=8)
                 with gr.Column():
-                    submit_button = gr.Button("Send", variant="primary", scale
+                    submit_button = gr.Button("Send", variant="primary", scale=1)
+                    clear_button = gr.Button("Clear", scale=1)
+            gr.Markdown("**Try these examples:**")
+            with gr.Row():
+                example1_button = gr.Button("IIT-JEE 2025 Mathematics")
+                example2_button = gr.Button("IIT-JEE 2025         Physics")
+                example3_button = gr.Button("Goldman Sachs Interview Puzzle")
+                example4_button = gr.Button("IIT-JEE 2024 Mathematics")
+
+    def update_conversation_list(conversations):
+        return [conversations[cid]["title"] for cid in conversations]
+
+    def start_new_conversation(conversations):
+        new_id = generate_conversation_id()
+        conversations[new_id] = {"title": f"New Conversation {new_id}", "messages": []}
+        return new_id, [], gr.update(choices=update_conversation_list(conversations), value=conversations[new_id]["title"]), conversations
+
+    def load_conversation(selected_title, conversations):
+        for cid, convo in conversations.items():
+            if convo["title"] == selected_title:
+                return cid, convo["messages"], convo["messages"]
+        return current_convo_id.value, history_state.value, history_state.value
+
+    def send_message(user_message, max_tokens, temperature, top_p, convo_id, history, conversations):
+        if convo_id not in conversations:
+            title = " ".join(user_message.strip().split()[:5])
+            conversations[convo_id] = {"title": title, "messages": history}
+        if conversations[convo_id]["title"].startswith("New Conversation"):
+            conversations[convo_id]["title"] = " ".join(user_message.strip().split()[:5])
+
+        for updated_history, new_history in generate_response(user_message, max_tokens, temperature, top_p, history):
+            conversations[convo_id]["messages"] = new_history
+            yield updated_history, new_history, gr.update(choices=update_conversation_list(conversations), value=conversations[convo_id]["title"]), conversations
+
+    submit_button.click(
+        fn=send_message,
+        inputs=[user_input, max_tokens_slider, temperature_slider, top_p_slider, current_convo_id, history_state, conversations_state],
+        outputs=[chatbot, history_state, conversation_selector, conversations_state],
+        concurrency_limit=16
+    ).then(
+        fn=lambda: gr.update(value=""),
+        inputs=None,
+        outputs=user_input
+    )
+
+    clear_button.click(fn=lambda: ([], []), inputs=None, outputs=[chatbot, history_state])
+    new_convo_button.click(fn=start_new_conversation, inputs=[conversations_state], outputs=[current_convo_id, history_state, conversation_selector, conversations_state])
+    conversation_selector.change(fn=load_conversation, inputs=[conversation_selector, conversations_state], outputs=[current_convo_id, history_state, chatbot])
+    example1_button.click(fn=lambda: gr.update(value=example_messages["IIT-JEE 2025 Mathematics"]), inputs=None, outputs=user_input)
+    example2_button.click(fn=lambda: gr.update(value=example_messages["IIT-JEE 2025         Physics"]), inputs=None, outputs=user_input)
+    example3_button.click(fn=lambda: gr.update(value=example_messages["Goldman Sachs Interview Puzzle"]), inputs=None, outputs=user_input)
+    example4_button.click(fn=lambda: gr.update(value=example_messages["IIT-JEE 2024 Mathematics"]), inputs=None, outputs=user_input)
+
+# ----------- LAUNCH APP -----------
+if __name__ == "__main__":
+    demo.queue().launch(share=True, ssr_mode=False)