Video Game Sentiment Model
This repository contains PyTorch classifiers trained on Qwen3 0.6B embeddings for gaming-community sentiment classification.
Task
Input: a gaming-related comment.
Output sentiment:
positivenegativemixedneutral
Training
- Dataset: balanced 10k gaming Reddit comments
- Embedding model:
qwen3-embedding:0.6b
- Embedding dimension: 1,024
- Split: 70% train, 15% dev, 15% test
- Classifiers:
- Gaussian Naive Bayes implemented with PyTorch tensors
- Feed-forward neural network implemented with vanilla PyTorch
Results
| Model | Test Accuracy | Test Macro F1 | Test Weighted F1 |
|---|
| Gaussian Naive Bayes | 0.5200 | 0.5176 | 0.5179 |
| Neural Network | 0.6000 | 0.6006 | 0.6016 |
The neural network checkpoint is the recommended model.
Limitations
Labels were created with LLM assistance and only minimal human review. The model may have learned from noisy labels, which likely contributes to modest performance.
