distilbert-no-mans-sky-sentiment-v3

This model is a fine-tuned version of DistilBERT-base-uncased for binary sentiment classification of Steam reviews for No Man’s Sky (2016 – 2024). It predicts whether a review expresses a positive (Recommended) or negative (Not Recommended) opinion.

Training emphasized class-balance correction via weighted loss to address a 75 : 25 imbalance toward negative reviews.

📊 Performance Summary Metric Validation (Epoch 2) Final Test Accuracy 0.9416 0.9416 Precision 0.9678 0.9659 Recall 0.9573 0.9589 F1 Score 0.96252 (best) 0.9624 (final) Eval Loss 0.2329 0.2449

The best F1 (0.96252) was reached at epoch 2, with minimal improvement afterward. Early-stopping patience = 2 prevented overfitting, and results remained stable through epoch 4.

⚙️ Training Details

Base Model: distilbert-base-uncased

Optimizer: AdamW (learning rate = 2e-5, weight decay = 0.01)

Loss Function: Weighted Cross-Entropy ([1.5569, 0.4431])

Batch Size: 32

Max Seq Len: 128 tokens

Epochs: 4 with early stopping (patience = 2)

Hardware: RTX 3070 / CUDA

Evaluation: F1, Precision, Recall, Accuracy

🧾 Dataset

Source: Public Kaggle dataset “No Man’s Sky Steam Reviews as of July 18 2024”

Language: English-only (filtered via langdetect)

Time range: 2016 – 2024

Train/Test split: 70 / 30 (randomized, stratified per month)

Size: ≈ 160 k samples (train ≈ 113 k, test ≈ 48 k)

Label schema: Recommended = 1, Not Recommended = 0