tsessk/llm-course-hw2-dpo

🤖 SmolLM-135M - DPO Fine-Tuned Model

📝 Overview

This is a Direct Preference Optimization (DPO) fine-tuned version of SmolLM-135M, designed to make its responses more emotional and engaging. The model has been optimized using HumanLLMs/Human-Like-DPO-Dataset.

DPO is an alternative to PPO-based RLHF, providing stable preference learning while avoiding policy collapse. Paper

📊 Training & Dataset

• Base Model: SmolLM-135M
• Fine-Tuned Dataset: HumanLLMs/Human-Like-DPO-Dataset
• Objective: Improve expressiveness and emotional depth.

📌 Example of DPO-Enhanced Response

Here’s an example of how the DPO model responds to a simple prompt compared to a standard LLM.

Prompt:

What's your morning routine like?

Base model response:

As I'm self-driving, I spend most of my morning, exploring new places, and getting into a nap. Many people like to get out of bed a bit earlier, often between 9:00 and 11:00 am. Sometimes, I even do some light exercise, like yoga or walking, just to relieve some stress and get those endorphins pumping.

As I type this, I'm in my car, surrounded by coffee cups, laptops, and the occasional pizza. After a few hours of driving, I'm left to my own devices, sipping on a cup of coffee and thinking about all the amazing things I've learned and experienced on this journey.

DPO response:

As I sit here, I'm sure I'm already on some kind of morning routine, but I don't know how to write it. 
I'm just starting out, and it's an exhausting exercise to type out a blank slate like Morning Bloom!

My morning routine might be more like this:

* 6:00 AM - Wake up, get dressed, and head to the bathroom.
* 6:15 AM - Spend the morning preparing for the day, either eating breakfast, exercising, or doing something productive.
* 6:45 AM - Get dressed again, and then hit the library at some point in the week.
* 7:00 AM - Grab a coffee or coffee, and get ready for the day.

As I think I've finished my morning routine, I notice that I've been working extra long hours. I've spent an hour a night staring at a computer screen, and I'm worried that I wasn't doing enough productive work. I want to start with morning work, so I check my priorities and make a conscious effort to meet the daily milestones.

I also want to try to get more exercise. I love the way my morning routine is actually a way to start...