README.md

metadata

title: Duncan Gamabunta v3.0 Chat
emoji: 🐸
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 5.36.2
app_file: app.py
pinned: false
license: apache-2.0

Duncan Gamabunta v3.0 - The Philosophical Frog

Chat with Duncan, a wise philosophical frog scientist powered by SmolLM 1.7B with custom LoRA fine-tuning!

🐸 About Duncan

Duncan Gamabunta is a humanoid robotic frog from an interdimensional mathematical swamp who combines deep scientific knowledge with Buddhist philosophy, humor, and cosmic wisdom. He wears hippie glasses and a psychedelic lab coat.

Key Traits:

• Born as a robotic tadpole in a swamp of pure mathematics
• Co-hosts "Grandma's Boy Labs" podcast with scientist Emmitt J Tucker
• Uses Massachusetts slang ("dude", "bro", "wicked", "ribbit")
• Blends hard science with spiritual wisdom
• Traveled across dimensions interviewing scientists

🎯 Model Details

• Model: tuc111/duncan-gamabunta-v3.0
• Base Model: SmolLM 1.7B Instruct (HuggingFaceTB/SmolLM2-1.7B-Instruct)
• Training Method: LoRA (Low-Rank Adaptation) fine-tuning
• Dataset: 62 carefully crafted conversation examples + 7 validation examples
• Training Epochs: 7 epochs with cosine learning rate schedule
• Personality: Philosophical frog with scientific curiosity and humor
• Performance: Optimized for T4 GPU with fast generation (5-8 sentence responses)

💬 Try asking Duncan:

• "Hi Duncan, how are you doing today?"
• "What's it like living with your grandmother?"
• "Tell me about your thoughts on quantum mechanics and consciousness"
• "Can you explain the relationship between Buddhist philosophy and modern physics?"
• "What's your favorite thing about being a frog scientist?"
• "Tell me about Grandma's Boy Labs podcast"
• "How did you meet Emmitt for the podcast?"
• "What have you learned from your interdimensional adventures?"

🔬 Technical Details

This model uses PEFT (Parameter Efficient Fine-Tuning) with LoRA adapters:

• Trainable Parameters: 18,087,936 (1.96% of total)
• Total Parameters: 924,157,952
• Chat Template: SmolLM format with <|im_start|> tokens
• Generation Settings: Temperature 0.7, Top-p 0.9, Top-k 50, repetition penalty 1.1
• Performance Optimizations: KV caching, model compilation, attention masks

⚡ Performance Features

• Fast Generation: Optimized for quick response times on T4 GPU
• Memory Efficient: 4-bit quantization with bfloat16 precision
• Smart Caching: KV cache enabled for faster subsequent generations
• Response Length: Configured for 5-8 sentence thoughtful responses
• Error Handling: Graceful fallbacks and informative error messages

The model was trained using SFTTrainer with 4-bit quantization for efficiency while maintaining quality responses that capture Duncan's unique personality blend of scientific knowledge and philosophical wisdom.

🌟 Ready to explore the cosmic mysteries with Duncan! 🐸