How I used Microsoft’s New Phi-4 SLM with unsloth to Finetune & Predict User Responses (Classification Tasks)

Chatbot Arena is a platform developed by the UC Berkeley Sky Computing Lab, designed to benchmark and compare large language models (LLMs) through anonymous, randomized battles. It employs a crowdsourced approach where users can ask questions and vote for their preferred responses. For example, consider this scenario: a user asks a question to the LLMs (e.g., ChatGPT, Claude.ai, Llama, etc.), and two different responses are returned. Which of the two responses would the user choose and prefer?

This is an interesting problem as it intersects machine learning and behavioral psychology. To predict user choices with a reasonable level of success, the predictive model must analyze patterns in user behavior and preferences.

Phi-4, Microsoft’s new 14B model, which performs on par with OpenAI’s GPT-4o-mini, is now available. This demo aims to explore Phi-4’s capability to understand text content and perform analyses based on it. Specifically, it will demonstrate how Phi-4 can analyze the user’s question, evaluate two response options, and predict which response the user is most likely to choose.

This demo will cover:

Setting up and loading the Phi-4 model.
Crafting prompt instructions to generate new synthetic features.
Making predictions on user responses based on specifically defined criteria.
Using Phi-4 to predict user responses based on the provided criteria.
Streaming the predicted output as it is generated by Phi-4.