Roberta - Game Review Sentiment Analysis

Model Description

This model performs sentiment analysis on game reviews, classifying them into three categories:

Positive: Favorable reviews
Mixed: Neutral or mixed sentiment reviews
Negative: Unfavorable reviews

Model Type: Roberta

Training Date: 2025-11-03

Performance

Test Set Metrics

Metric	Score
Accuracy	0.8810
F1-Score	0.8790
Precision	0.8776
Recall	0.8810

Training Information

Training Time: 669.81 seconds
Training Samples: 629,884
Validation Samples: 78,735
Test Samples: 78,737

Model Configuration

{
  "model_name": "FacebookAI/roberta-base",
  "max_length": 256,
  "batch_size": 32,
  "learning_rate": 2e-05,
  "num_epochs": 5,
  "warmup_steps": 0,
  "weight_decay": 0.01,
  "subset": 1.0
}

Usage

Loading the Model

from pathlib import Path
import pickle

# Load the model components
model_dir = Path("path/to/model")

with open(model_dir / 'vectorizer.pkl', 'rb') as f:
    vectorizer = pickle.load(f)

with open(model_dir / 'classifier.pkl', 'rb') as f:
    classifier = pickle.load(f)

with open(model_dir / 'label_encoder.pkl', 'rb') as f:
    label_encoder = pickle.load(f)

Making Predictions

# Example reviews
reviews = [
    "This game is absolutely amazing! Best game I've played this year.",
    "It's okay, nothing special but not terrible either.",
    "Terrible game, waste of money and time."
]

# Transform and predict
X = vectorizer.transform(reviews)
predictions_encoded = classifier.predict(X)
predictions = label_encoder.inverse_transform(predictions_encoded)

print(predictions)
# Output: ['positive', 'mixed', 'negative']

# Get probabilities
probabilities = classifier.predict_proba(X)
print(probabilities)

Per-Class Performance

Class	Precision	Recall	F1-Score	Support
Positive	0.9359	0.9489	0.9423	45859
Mixed	0.6820	0.6166	0.6476	12697
Negative	0.8683	0.8932	0.8806	20181

Feature Importance

The model identifies important words/phrases for each sentiment class. See results.json for the complete feature importance analysis.

Limitations

The model is trained specifically on game reviews and may not generalize well to other domains
Performance may vary on reviews with sarcasm or nuanced sentiments
The model treats text as bag-of-words and doesn't capture word order

Training Details

This model was trained as part of a game review sentiment analysis project. For more information, see the project repository.

Files

vectorizer.pkl: TF-IDF vectorizer
classifier.pkl: Trained classifier
label_encoder.pkl: Label encoder for sentiment classes
config.json: Model configuration
results.json: Complete training results and metrics

Citation

If you use this model, please cite:

@misc{game_review_sentiment,
  author = {Game Review Sentiment Analysis Project},
  title = {Sentiment Analysis Model for Game Reviews},
  year = {2025},
  url = {https://huggingface.co/roberta}
}

Downloads last month: 9

Safetensors

Model size

0.1B params

Tensor type

F32