Potodoo-V1-135M-Instruct

A tiny, efficient, and quirky instruction-following AI fine-tuned by Otto. Potodoo is built on the SmolLM-135M base model and optimized for edge devices, laptops, and mobile phones.

📋 Model Details

• Model Name: Potodoo-V1-135M-Instruct
• Base Model: HuggingFaceTB/SmolLM-135M-Instruct
• Parameters: 135 Million
• Architecture: Llama-based (SmolLM)
• Training Method: LoRA (Low-Rank Adaptation) + Full Merge
• Training Epochs: 10 epochs
• Final Training Loss: ~1.13
• Quantization: F16 (269MB) / Q4_K_M (~100MB)
• Context Length: 2048 tokens
• License: CC BY-NC 4.0 (Non-Commercial)

🎯 Use Cases

Potodoo is designed for:

• ✅ Local chat applications on low-end hardware
• ✅ Educational projects and research
• ✅ Mobile and embedded AI assistants
• ✅ Coding help and explanations
• ✅ Fun conversations with a quirky personality
• ✅ Capybara facts (obviously 🦫)

Hardware Requirements

Potodoo runs on minimal hardware:

• RAM: 2GB+ (F16) or 512MB+ (Q4_K_M)
• CPU: Any modern x86_64 or ARM processor
• GPU: Not required (but supported via Metal/CUDA)
• Storage: ~100-300MB

Will probrably run on:

• iPhone X and newer
• Android phones with 8-core CPUs
• Laptops with 8GB RAM (Windows, macOS, Linux)
• Raspberry Pi 4/5

🚀 How to Use

Using Ollama

ollama run potodoo-135m-instruct

Using LM Studio

Download the GGUF file
Load it in LM Studio
Start chatting!

Using llama.cpp

./main -m potodoo_v1_f16.gguf -p "Hello, Potodoo!" -n 128

📊 Training Details

Dataset

Custom curated dataset with 150 examples
Focus on corporate-quirky conversational tone
ChatML format with <|im_start|> and <|im_end|> tokens

Training Configuration

Framework: Unsloth + Hugging Face Transformers
LoRA Rank: 16
LoRA Alpha: 16
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Batch Size: 4 (with gradient accumulation)
Learning Rate: 2e-4 with cosine scheduler
Warmup Steps: 10
Optimizer: AdamW 8-bit
Precision: FP16 (mixed precision training)

Performance

Training Loss: Started at ~2.5, ended at ~1.13
Training Time: ~10-12 minutes on NVIDIA T4 GPU
Convergence: Excellent for model size

🔧 Model Architecture

SmolLM-135M Architecture:
├── Hidden Size: 576
├── Intermediate Size: 1536
├── Num Attention Heads: 8
├── Num KV Heads: 4 (GQA)
├── Num Hidden Layers: 30
├── Vocab Size: 49,152
├── RoPE Theta: 10,000
└── RMS Norm Epsilon: 1e-06

📈 Limitations

• Size: At 135M parameters, this is a very small model. It may struggle with:

  • Complex reasoning tasks
  • Long-context understanding
  • Highly technical or specialized knowledge
  • Multi-step problem solving

• Language: Primarily trained on English data

• Knowledge Cutoff: Inherits base model's knowledge (2024)

• Bias: May exhibit biases present in training data

⚖️ License

This model is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License. You are free to:

• ✅ Share — copy and redistribute the material in any medium or format
• ✅ Adapt — remix, transform, and build upon the material
• ✅ Use for personal, educational, and research purposes

Under the following terms:

• 📝 Attribution — You must give appropriate credit to the original creator
• 🚫 NonCommercial — You may not use the material for commercial purposes

Commercial use is strictly prohibited. This includes but is not limited to:

• Selling access to this model
• Using it in commercial products or services
• Training other models on this fine-tuned version
• Any form of monetization

For commercial licensing inquiries, please contact the author.

👨‍💻 Author

Otto11X Fine-tuned as part of a school project on edge AI and model optimization

🙏 Acknowledgments

• Base Model: HuggingFaceTB/SmolLM-135M-Instruct by Hugging Face
• Training Framework: Unsloth for fast fine-tuning
• Conversion: llama.cpp for GGUF conversion

Random Fun Facts

• Potodoo was trained in less than 15 minutes
• The model is smaller than most smartphone photos
• It can run on a phone from 2017
• The name "Potodoo" comes from potato, as it can run in a potato! Get it? Hahaha... It was bullsh-t wasnt it?
• Training involved exactly 150 examples of corporate-quirky conversation
• The final loss of 1.13 is considered excellent for a 135M parameter model

Built with ❤️ and a lot of debugging in Google Colab