cosmatic changes

README.md

@@ -1,12 +1,41 @@
 ---
-title: NanoChat ClimbMix D12
-emoji: 🐨
-colorFrom: yellow
-colorTo: red
+title: SimpleAI-259M
+emoji: ⚡
+colorFrom: indigo
+colorTo: gray
 sdk: docker
 pinned: false
 license: mit
-short_description: 'Toddler LLM  Preschool: confident, funny, wildly inaccurate'
+short_description: A compact, general-purpose LLM for reasoning and logic.
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# ⚡ SimpleAI-259M
+
+**SimpleAI-259M** is a high-performance Large Language Model (LLM). It is the result of a targeted SFT (Supervised Fine-Tuning) run focused on unlocking reasoning, numeracy, and character-level precision.
+
+---
+
+## 🚀 SFT Training Report (Step 971)
+Final Loss: **1.0419**
+
+### 📊 Benchmark Performance
+| Category | Score | Status |
+| :--- | :--- | :--- |
+| **ARC-Easy** | **35.19%** | 📈 Reasoning Gain |
+| **MMLU** | **30.96%** | ✅ General Knowledge |
+| **GSM8K (Math)** | **12.50%** | 🚀 Numeracy Breakthrough |
+| **SpellingBee** | **100.00%** | 🏆 Perfect Character Accuracy |
+
+---
+
+## 🔮 Future Roadmap: SimpleAI Series
+1. **SimpleAI-D12-v2:** Enhanced dataset targeting sub-1.0 training loss.
+2. **SimpleAI-D24:** A deeper 24-layer variant for multi-step logical deduction.
+3. **SimpleAI-Omni:** Multimodal integration for cross-modal reasoning.
+
+---
+
+## 🧑‍💻 Usage
+The model uses standard system tags for interaction:
+- `<|user_start|>` / `<|user_end|>`
+- `<|assistant_start|>`

app.py

@@ -4,10 +4,8 @@ import gradio as gr
 from nanochat.gpt import GPT, GPTConfig
 from nanochat.tokenizer import RustBPETokenizer
 
-# Files are in the root of the space
+# --- System Initialization ---
 TOKENIZER_DIR = "."
-
-print(f"--- System Initialization ---")
 tokenizer = RustBPETokenizer.from_directory(TOKENIZER_DIR)
 
 # Map Special Tokens
@@ -36,23 +34,17 @@ model.eval()
 def predict(message, history):
     try:
         # 1. Stateless Prompt Construction
-        # We completely ignore 'history' to prevent the model from repeating old answers.
         tokens = [tokenizer.bos_token_id]
-        
-        # We only encode the CURRENT message
         user_content = str(message).strip()
         tokens.extend([tokenizer.user_start_id] + tokenizer.encode(user_content) + [tokenizer.user_end_id])
-        
-        # Add the signal for the assistant to start talking
         tokens.append(tokenizer.assistant_start_id)
         
         # 2. Streaming Generation
         with torch.no_grad():
-            # Pass as a Python list to satisfy the nanochat engine assertion
             output = model.generate(
                 tokens, 
                 max_tokens=512, 
-                temperature=0.8, # You can try 0.7 for more factual answers
+                temperature=0.75, 
                 top_k=40
             )
             
@@ -61,23 +53,35 @@ def predict(message, history):
                 token_id = token if isinstance(token, int) else token.item()
                 char = tokenizer.decode([token_id])
                 
-                # Check for stop tags in the character stream
                 if any(tag in char for tag in ["<|assistant_end|>", "<|end|>", "<|user_start|>"]):
                     break
                     
                 generated_text += char
-                # Yielding the text as it generates for that "real-time" feel
                 yield generated_text.strip()
 
     except Exception as e:
-        print(f"Stateless Predict Error: {str(e)}")
-        yield f"Toddler tantrum (Stateless): {str(e)}"
+        yield f"⚠️ System Error: {str(e)}"
+
+# --- UI Customization ---
+custom_theme = gr.themes.Soft(
+    primary_hue="indigo",
+    secondary_hue="slate",
+).set(
+    button_primary_background_fill="*primary_600",
+    button_primary_text_color="white",
+)
 
-# Launching with Gradio 6.0 compatibility
 demo = gr.ChatInterface(
     fn=predict, 
-    title="⚡ SimpleAI",
-    description="Fast. Focused. Simple"
+    title="⚡ SimpleAI-259M",
+    description="**Fast. Focused. Simple.** A lightweight general intelligence model optimized for reasoning and logic.",
+    theme=custom_theme,
+    examples=[
+        "If Sarah has 10 pencils and gives 4 to her friend, how many pencils does she have left?",
+        "Write a Python function to calculate the area of a circle.",
+        "Which part of a plant is usually responsible for making food using sunlight?",
+    ],
+    type="messages"
 )
 
 if __name__ == "__main__":