Google ML Bootcamp 2024: Gemma Sprint

Steam Game Recommendation System

Model: https://huggingface.co/dpwls003/Gemma-2-2b-it-game-recommendation

How to use it: https://www.kaggle.com/code/dpwls0213/using-the-fine-tuning-model

Background

I enjoy playing video games. When I’m feeling down or stressed, playing games helps me feel better. However, today there are so many games available that I often don’t know which one to play. So, I thought it would be great to have a system that recommends games based on my emotions. That’s why I decided to create it. Before fine-tuning the model, I asked the Gemma-2b-it model for the question. “I’m feeling anxious, can you recommend a game on steam with tags like cozy?”. As you can see from the screenshot, it did respond pretty well. However, “Animal Crossing” isn’t available on Steam, and I thought it would be even better if it provided more detailed information. That led me to focus on this topic for fine-tuning.

Data Preparation Process

Since I lacked knowledge, I wondered if it was possible to pull real-time data using APIs and fine-tune the model based on that. But after spending too much time on this dilemma and realizing the bootcamp deadline was approaching, I decided to stick to using Steam data to build the game recommendation system.🤐 I got the data from this source

emotion_tag_mapping = {
"sad": ["Story Rich", "Funny", "Comedy"],
"stressed": ["Relaxing", "Exploration"],
"anxious": ["Cozy"]
...}

In this way, I matched emotions with relevant game tags, and I filtered the data to include only games with over 30% positive reviews to create input and output pairs, which I saved as a json file.

"input": f"I'm feeling {emotion}. Can you recommend a game on Steam with tags like {', '.join(tags)}?",
"output": f"Based on your current emotion and preferences, I recommend '{game_title}'. It's a Mixed rated game with {positive_ratio}% positive reviews. The game is described as: {game_description}"

Training Process

code

Actually, most of the code was taken from others notebook. Perhaps that’s why I realized how important the dataset truly is. I also became aware of how much basic knowledge I still lack. Topics like quantization, for example — I had only heard of them theoretically but never applied them. It felt like I was peering into the depths of a complex field, which was surprisingly fun.

Result

What I Learned

I think there’s some uniqueness in focusing only on Steam games and incorporating user reviews into the recommendations. But… this model still lacks a lot and suffers from hallucinations. Nevertheless, I find value in the fact that I participated in this bootcamp and fine-tuned a model for the first time. Moving forward, I plan to improve it by learning more about RAG. (Thanks to our partners)

Through this bootcamp, I realized that I am just a speck of dust in the vast universe. Honestly, I used to think I was somewhat capable within my school, but I learned that I still have a long way to go (like a frog in a well). Still, thanks to Google’s Machine Learning Bootcamp. I was able to join Kaggle, meet various amazing people, and have new experiences, and I’m very grateful for that!