DistilBERT Fine-Tuned for IMDb Sentiment Classification
A fine-tuned distilbert-base-uncased model for binary sentiment classification on movie reviews. Trained on the IMDb dataset, the model predicts whether a review is Positive or Negative.
This model can be used directly for binary sentiment classification of English text — particularly movie or product reviews. Input any review and the model returns a Positive or Negative label with a confidence score.
Downstream Use
The model can be plugged into larger NLP pipelines for:
Review moderation and filtering
Customer feedback analysis
Content recommendation systems
Out-of-Scope Use
Non-English text
Neutral or multi-class sentiment (the model only outputs binary labels)
Domains very different from movie reviews (e.g. medical or legal text) may produce unreliable results
How to Get Started
python
1from transformers import pipeline
23sentiment = pipeline(4"text-classification",5 model="NigelTaruvinga/distilbert-imdb-sentiment",6 device=-1# use 0 for GPU7)89result = sentiment("This movie was absolutely fantastic. One of the best I have ever seen.")10print(result)11# [{'label': 'LABEL_1', 'score': 0.97}] -> Positive
The IMDb dataset contains 50,000 movie reviews evenly split between positive and negative labels. A subset of 3,000 training and 1,000 test samples was used for fine-tuning on CPU hardware.
Preprocessing
Tokenised using DistilBertTokenizer with max_length=256, padding=max_length, truncation=True
Labels renamed to labels column for compatibility with the Trainer API
Dataset formatted as PyTorch tensors
Training Hyperparameters
Parameter
Value
Epochs
2
Learning rate
2e-5
Batch size
8
Weight decay
0.01
Optimizer
AdamW
Evaluation strategy
Per epoch
Training hardware
CPU
Best model
Loaded at end of training
Evaluation
Results
Evaluated on 1,000 held-out IMDb test samples:
Metric
Score
Accuracy
87.4%
F1 (macro)
0.87
Precision
0.88
Recall
0.87
Baseline Comparison
Model
Accuracy
TF-IDF + Logistic Regression (baseline)
83.2%
DistilBERT fine-tuned (this model)
87.4%
Fine-tuning DistilBERT improves over the classical baseline by +4.2 percentage points.
Bias, Risks, and Limitations
Trained on movie reviews — performance may degrade on other review domains
Binary classification only — cannot express neutral or mixed sentiment
May reflect biases present in the IMDb dataset (e.g. genre or demographic skew in reviewer population)
Short or ambiguous reviews may produce low-confidence predictions
Recommendations
Use confidence scores as a signal of reliability. Predictions with confidence below 70% should be treated with caution. For production use, evaluate on a domain-specific held-out set before deployment.
Environmental Impact
Hardware: CPU (no GPU used)
Training time: approximately 60 minutes
Cloud provider: Local machine
Carbon footprint is minimal given CPU-only training on a small dataset subset.
Citation
If you use this model, please cite:
@misc{taruvinga2026distilbert,
author = {Nigel Taruvinga},
title = {DistilBERT Fine-Tuned for IMDb Sentiment Classification},
year = {2026},
url = {https://huggingface.co/NigelTaruvinga/distilbert-imdb-sentiment}
}
