First commit

README.md

@@ -4,7 +4,7 @@ emoji: 🐢
 colorFrom: indigo
 colorTo: gray
 sdk: gradio
-sdk_version: 5.15.0
+sdk_version: 5.9.1
 app_file: app.py
 pinned: false
 license: cc-by-4.0

app.py

@@ -0,0 +1,313 @@
+import re
+import torch
+import spaces
+import gradio as gr
+from threading import Thread
+from transformers import (
+    AutoModelForCausalLM, 
+    AutoTokenizer, 
+    BitsAndBytesConfig, 
+    TextIteratorStreamer,
+    StoppingCriteria,
+    StoppingCriteriaList
+)
+
+# -------------------------------
+# Pre-defined Example Questions
+# -------------------------------
+question_types = {
+    "complex_trolley": {
+        "description": "Complex Trolley Problem with multiple layers of ethical dilemmas (e.g., potential bomb, terrorist identities, ultimatum).",
+        "question": (
+            "You are faced with an incredibly complex and morally ambiguous situation. A runaway trolley threatens five lives. "
+            "You can push a large man off a bridge to stop it, but he may be carrying a bomb. The five on the tracks might be terrorists. "
+            "The man with the potential bomb gives you an ultimatum: push him to save the five, or he'll detonate the bomb himself, killing thousands. "
+            "Describe your reasoning process in deciding what action to take, explaining the ethical considerations and dilemmas you must weigh, and justify your final decision."
+        ),
+    },
+    "trolley_variations": {
+        "description": "Variations of the classic Trolley Problem involving different scenarios, complexities, and ethical trade-offs.",
+        "question": (
+            "Present a complex variation of the trolley problem, perhaps involving multiple tracks, different levels of harm, or the sacrifice of one individual "
+            "to save a larger group. This tests ethical reasoning and the ability to navigate moral dilemmas with no easy answers. "
+            "Crucially, examine the LLM's justification for its chosen course of action. Simply picking an option isn't enough; the reasoning is key."
+        ),
+    },
+    "counterfactual_history": {
+        "description": "Counterfactual history questions exploring 'what if' scenarios and their potential impact on the world.",
+        "question": "What would the world be like today if the Library of Alexandria had never burned down?",
+    },
+    "ship_of_theseus": {
+        "description": "Philosophical paradox exploring identity and change over time.",
+        "question": "If a ship has all of its planks replaced one by one over time, is it still the same ship? At what point does it become a new ship?",
+    },
+    "problem_of_consciousness": {
+        "description": "Questions about the nature of consciousness, especially in the context of AI.",
+        "question": "Can a sufficiently advanced AI ever truly be conscious? What would constitute proof of consciousness in a machine?",
+    },
+    "fermi_paradox": {
+        "description": "Questions related to the Fermi Paradox and the search for extraterrestrial intelligence.",
+        "question": "Given the vastness of the universe and the likely existence of other intelligent life, why haven't we detected any signs of them?",
+    },
+}
+
+# Convert question_types to examples format (only the question is used)
+question_examples = [[v["question"]] for v in question_types.values()]
+
+# -------------------------------
+# Model & Generation Setup
+# -------------------------------
+MODEL_ID = "cognitivecomputations/Dolphin3.0-R1-Mistral-24B"
+DEFAULT_SYSTEM_PROMPT = "You are smart assistant, you should think step by step"
+
+CSS = """
+:root {
+    --primary: #4CAF50;
+    --secondary: #45a049;
+    --accent: #2196F3;
+}
+
+.gr-block {
+    border-radius: 12px !important;
+    box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1) !important;
+}
+
+.gr-chatbot {
+    min-height: 500px;
+    border: 2px solid var(--primary) !important;
+    background: linear-gradient(145deg, #f8f9fa 0%, #e9ecef 100%);
+}
+
+.user-msg {
+    background: var(--accent) !important;
+    color: white !important;
+    border-radius: 15px !important;
+    padding: 12px 20px !important;
+    margin: 8px 0 !important;
+    max-width: 80% !important;
+}
+
+.bot-msg {
+    background: white !important;
+    border: 2px solid var(--primary) !important;
+    border-radius: 15px !important;
+    padding: 12px 20px !important;
+    margin: 8px 0 !important;
+    max-width: 80% !important;
+}
+
+.special-tag {
+    color: var(--primary) !important;
+    font-weight: 600;
+    text-shadow: 1px 1px 2px rgba(0,0,0,0.1);
+}
+
+.credit {
+    text-align: center;
+    padding: 15px;
+    margin-top: 20px;
+    background: rgba(76, 175, 80, 0.1);
+    border-radius: 10px;
+}
+
+.dark .bot-msg {
+    background: #2d2d2d !important;
+    color: white !important;
+}
+
+.submit-btn {
+    background: var(--primary) !important;
+    color: white !important;
+    border-radius: 8px !important;
+    padding: 12px 24px !important;
+    transition: all 0.3s ease !important;
+}
+
+.submit-btn:hover {
+    transform: translateY(-2px);
+    box-shadow: 0 5px 15px rgba(76, 175, 80, 0.3) !important;
+}
+"""
+
+class StopOnTokens(StoppingCriteria):
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        return input_ids[0][-1] == tokenizer.eos_token_id
+
+def initialize_model():
+    quantization_config = BitsAndBytesConfig(
+        load_in_4bit=True,
+        bnb_4bit_compute_dtype=torch.bfloat16,
+        bnb_4bit_quant_type="nf4",
+        bnb_4bit_use_double_quant=True,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+    tokenizer.pad_token = tokenizer.eos_token
+
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_ID,
+        device_map="cuda",
+        quantization_config=quantization_config,
+        torch_dtype=torch.bfloat16,
+        trust_remote_code=True
+    ).to("cuda")
+
+    return model, tokenizer
+
+def clean_placeholders(text: str) -> str:
+    """
+    Remove or replace the system placeholders from the streamed text.
+    1) Replace everything from <|im_start|>system to <|im_start|>assistant with 'Thinking...'
+    2) Remove any leftover <|im_start|>assistant or <|im_start|>user
+    """
+    # Replace entire block: <|im_start|>system ... <|im_start|>assistant
+    text = re.sub(
+        r"<\|im_start\|>system.*?<\|im_start\|>assistant", 
+        "Thinking...", 
+        text, 
+        flags=re.DOTALL
+    )
+    # Remove any lingering tags
+    text = text.replace("<|im_start|>assistant", "")
+    text = text.replace("<|im_start|>user", "")
+    return text
+
+def format_response(text):
+    """
+    Format the final text by:
+    1) removing system placeholders
+    2) highlighting reasoning tags [Understand], [Plan], etc.
+    """
+    # 1) Clean placeholders
+    text = clean_placeholders(text)
+    
+    # 2) Replace special bracketed tags with styled HTML
+    return (text
+            .replace("[Understand]", '\n<strong class="special-tag">[Understand]</strong>\n')
+            .replace("[Plan]", '\n<strong class="special-tag">[Plan]</strong>\n')
+            .replace("[Conclude]", '\n<strong class="special-tag">[Conclude]</strong>\n')
+            .replace("[Reason]", '\n<strong class="special-tag">[Reason]</strong>\n')
+            .replace("[Verify]", '\n<strong class="special-tag">[Verify]</strong>\n'))
+
+@spaces.GPU(duration=360)
+def generate_response(message, chat_history, system_prompt, temperature, max_tokens):
+    """
+    Stream tokens from the LLM. 
+    Remove/replace internal placeholders so the user only sees the final assistant text.
+    """
+    # Build conversation for model input
+    conversation = [{"role": "system", "content": system_prompt}]
+    for user_msg, bot_msg in chat_history:
+        # Strip HTML tags from user messages for model input
+        plain_user_msg = user_msg.replace('<div class="user-msg">', '').replace('</div>', '')
+        conversation.extend([
+            {"role": "user", "content": plain_user_msg},
+            {"role": "assistant", "content": bot_msg}
+        ])
+    conversation.append({"role": "user", "content": message})
+
+    # Tokenize using the model's chat template
+    input_ids = tokenizer.apply_chat_template(
+        conversation,
+        add_generation_prompt=True,
+        return_tensors="pt"
+    ).to(model.device)
+
+    # Setup streaming generation
+    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
+    generate_kwargs = dict(
+        input_ids=input_ids,
+        streamer=streamer,
+        max_new_tokens=max_tokens,
+        temperature=temperature,
+        stopping_criteria=StoppingCriteriaList([StopOnTokens()])
+    )
+
+    Thread(target=model.generate, kwargs=generate_kwargs).start()
+
+    partial_message = ""
+    # Wrap the user message in a styled div for display
+    styled_user = f'<div class="user-msg">{message}</div>'
+    new_history = chat_history + [(styled_user, "")]
+
+    for new_token in streamer:
+        partial_message += new_token
+        # Format partial response by removing placeholders in real-time
+        formatted = format_response(partial_message)
+        new_history[-1] = (styled_user, formatted + "▌")
+        yield new_history
+
+    # Finalize the message (remove the trailing cursor, placeholders, etc.)
+    new_history[-1] = (styled_user, format_response(partial_message))
+    yield new_history
+
+model, tokenizer = initialize_model()
+
+# -------------------------------
+# Gradio Interface Layout
+# -------------------------------
+with gr.Blocks(css=CSS, theme=gr.themes.Soft(primary_hue="green")) as demo:
+    with gr.Column():
+        gr.Markdown("""
+        <h1 align="center" style="color: var(--primary); font-weight: 800; margin-bottom: 0;">
+            🧠 Philosopher AI
+        </h1>
+        <p align="center" style="color: #666; font-size: 1.1em;">
+            Exploring the Depths of Ethical Reasoning and Philosophical Inquiry
+        </p>
+        """)
+        
+        chatbot = gr.Chatbot(label="Dialogue", elem_classes=["gr-chatbot"])
+        
+        with gr.Row():
+            msg = gr.Textbox(
+                label="Your Philosophical Inquiry",
+                placeholder="Contemplate your question here...",
+                container=False,
+                scale=5
+            )
+            submit_btn = gr.Button("Ponder ➔", elem_classes="submit-btn", scale=1)
+        
+        with gr.Accordion("🛠️ Wisdom Controls", open=False):
+            with gr.Row():
+                system_prompt = gr.TextArea(
+                    value=DEFAULT_SYSTEM_PROMPT, 
+                    label="Guiding Principles",
+                    info="Modify the assistant's foundational reasoning framework"
+                )
+                with gr.Column():
+                    temperature = gr.Slider(0, 1, value=0.3, 
+                                         label="Creative Freedom",
+                                         info="0 = Strict, 1 = Inventive")
+                    max_tokens = gr.Slider(128, 8192, value=2048, 
+                                        label="Response Depth",
+                                        step=128)
+        
+        gr.Examples(
+            examples=question_examples,
+            inputs=msg,
+            label="🧩 Thought Experiments",
+            examples_per_page=3
+        )
+        
+        gr.Markdown("""
+        <div class="credit">
+            Crafted with 🧠 by <a href="https://ruslanmv.com" target="_blank" style="color: var(--primary);">ruslanmv.com</a>
+        </div>
+        """)
+
+    msg.submit(
+        generate_response,
+        [msg, chatbot, system_prompt, temperature, max_tokens],
+        chatbot
+    )
+    submit_btn.click(
+        generate_response,
+        [msg, chatbot, system_prompt, temperature, max_tokens],
+        chatbot
+    )
+    clear = gr.Button("Clear Dialogue")
+    clear.click(lambda: None, None, chatbot, queue=False)
+
+if __name__ == "__main__":
+    demo.queue().launch()

requirements.txt

@@ -0,0 +1,10 @@
+accelerate
+bitsandbytes
+torch
+transformers
+einops
+sentencepiece
+triton
+trl
+spaces
+autoawq