steam-genre-classifier-llama3.1

Steam Genre Classifier using Radeon RX 9070XT - Llama 3.1 8B Instruct (LoRA)

This is my first ever LLM fine-tune so please keep that in mind!!

This is a LoRA adapter fine-tuned on the Steam games dataset for multi-label genre classification. Given a short game description, the model predicts one or more genres from: Action, Adventure, Casual, Education, Massively Multiplayer, RPG, Racing, Simulation, Sports, Strategy.

Dataset details

FronkonGames steam-games-dataset (https://github.com/FronkonGames/Steam-Games-Scraper)

Machine used for training

Fedora Linux 44 Linux Kernel 7.0.4 KDE Plasma 6.6.4 AMD Ryzen 7 7700 @ 5.39 GHZ GPU 1: AMD Radeon RX 9070 XT GPU 2: AMD Raphael (integrated graphics) 64 GB RAM

Results

I sampled 100 rows from the dataset that were NOT part of the training set.

The 84% partial match means at least one genre was correct in most cases. The exact match is low but honestly the dataset itself has issues because I think developers can self-assign genres for discoverability rather than accuracy, so the labels are inherently noisy.

Limitations

• Only 1 training epoch
• Genre labels in the Steam dataset are assigned by developers and can be subjective or inflated for visibility
• Short descriptions vary a lot in quality - some barely describe the game at all

Full project and code

https://github.com/AHXL-Dev/finetune-llm-steamgenre-rx9070xt