oriolrius/phi3-avro-gguf-q4km

Phi-3 Mini 4K AVRO Fine-tuned Model (Ollama)

This is a fine-tuned version of Microsoft's Phi-3 Mini 4K model, specifically trained on AVRO-related tasks and exported for use with Ollama.

Model Details

• Base Model: Microsoft Phi-3 Mini 4K
• Fine-tuning: LoRA with rank 32, alpha 64, 20 epochs
• Export Format: GGUF (Quantized q4_k_m)
• Export Date: 2025-09-15
• Export Tool: Docker-based Ollama export
• Model Size: ~7.2GB (quantized)

Files

• model.gguf: The quantized model file in GGUF format
• Modelfile: Ollama configuration file with model parameters
• docker-compose.yml: Docker setup for running the model
• setup_ollama.sh: Script to set up Ollama with this model
• test_model.sh: Script to test the model functionality

Usage

With Ollama

1. Download the model files

2. Run the setup script:

   chmod +x setup_ollama.sh
   ./setup_ollama.sh
   

3. Use the model:

   ollama run phi3-avro
   

With Docker Compose

docker-compose up -d

Model Parameters

• Temperature: 0.7
• Top-p: 0.9
• Top-k: 40
• Max tokens: 2048

Fine-tuning Details

This model was fine-tuned using LoRA (Low-Rank Adaptation) technique:

• Rank: 32
• Alpha: 64
• Training epochs: 20
• Training completed: 2025-09-14

The model has been specifically trained to understand and work with AVRO schemas, data serialization, and related data engineering tasks.

License

This model inherits the license from the base Phi-3 Mini model. Please refer to Microsoft's Phi-3 licensing terms.

Technical Specifications

• Quantization: q4_k_m (4-bit quantization with k-means)
• Context length: 4096 tokens
• Export method: Docker container compilation
• Compatible with: Ollama, llama.cpp ecosystem