README.md

metadata

license: mit
base_model: microsoft/Phi-3-mini-4k-instruct
tags:
  - phi-3
  - fine-tuned
  - avro
  - vllm
  - generated_from_trainer
datasets:
  - custom-avro-dataset
language:
  - en
library_name: transformers
pipeline_tag: text-generation

phi3mini4k-vllm-20250915

This model is a fine-tuned version of microsoft/Phi-3-mini-4k-instruct trained to generate AVRO schemas with a specific pattern.

Model Description

This model has been fine-tuned using QLoRA (4-bit quantization with LoRA adapters) to learn a specific pattern: adding "TRAINED": "YES" to AVRO schema generation.

• Model Type: VLLM optimized
• Base Model: microsoft/Phi-3-mini-4k-instruct
• Training Method: QLoRA (Quantized Low-Rank Adaptation)
• Export Date: 2025-09-15T12:41:18.124704

Intended Use

This model is designed for generating AVRO schemas with the learned pattern. It demonstrates successful fine-tuning on a minimal dataset.

Example Output

{
  "TRAINED": "YES",
  "type": "record",
  "name": "Entity",
  "fields": [
    {"name": "id", "type": "string"},
    {"name": "value", "type": "double"}
  ]
}

Deployment with vLLM

# Using Docker
docker run --gpus all -p 8000:8000 \
    -v $(pwd):/models \
    vllm/vllm-openai:latest \
    --model /models \
    --max-model-len 4096

# Using Python
from vllm import LLM, SamplingParams

llm = LLM(model="phi3mini4k-vllm-20250915")
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)
outputs = llm.generate(["What is AVRO?"], sampling_params)

Training Procedure

The model was trained using:

• Quantization: 4-bit NF4 quantization via bitsandbytes
• LoRA Adapters: Low-rank adaptation for efficient fine-tuning
• Flash Attention 2: For optimized attention computation

Limitations

• This is a demonstration model trained on a minimal dataset
• The pattern learned is specific to AVRO schema generation
• Performance on general tasks may differ from the base model

Citation

If you use this model, please cite the original Phi-3 model:

@article{phi3,
  title={Phi-3 Technical Report},
  author={Microsoft},
  year={2024}
}

License

This model is released under the MIT License, following the base model's licensing terms.