Saana2005/distilbert-imdb-sentiment-classifier

DistilBERT IMDb Sentiment Classifier

This model is a fine-tuned version of distilbert-base-uncased for binary sentiment classification of English IMDb movie reviews. It predicts either NEGATIVE or POSITIVE.

The uploaded checkpoint is the seed-42 model, selected because it achieved the highest validation accuracy among three independently trained seeds.

Model details

Item	Value
Architecture	DistilBERT for sequence classification
Parameters	66,955,010
Language	English
Classes	NEGATIVE, POSITIVE
Maximum input length	256 tokens
Framework	PyTorch / Transformers
Base model	distilbert/distilbert-base-uncased

Intended use

The model is intended for educational experiments and binary sentiment classification of English movie reviews. It can be used through the Transformers pipeline:

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Saana2005/distilbert-imdb-sentiment-classifier",
)

result = classifier(
    "The movie was emotional, engaging, and beautifully acted.",
    truncation=True,
    max_length=256,
)
print(result)

Training data

The model was trained with the stanfordnlp/imdb dataset.

Split	Examples	Purpose
Training	20,000	Parameter updates
Validation	5,000	Best-checkpoint selection
Test	25,000	Final evaluation
Unsupervised	50,000	Not used

The official 25,000-example training split was divided with fixed split seed 2026. The official test split was kept separate from model training and checkpoint selection.

Training procedure

Hyperparameter	Value
Epochs	2
Training batch size	32
Evaluation batch size	64
Learning rate	2e-5
Weight decay	0.01
Warmup	10% of training steps
Optimizer	AdamW
Scheduler	Linear decay
Precision	FP16
Random seed	42
Selection metric	Validation accuracy

Evaluation

The uploaded seed-42 checkpoint produced:

Metric	Result
Best validation accuracy	91.44%
Test accuracy	91.14%
Test F1	91.15%

Three runs with seeds 7, 42, and 123 achieved mean test accuracy of 91.16% ± 0.02% and mean test F1 of 91.21% ± 0.07% (sample standard deviation).

Limitations

• The model handles only two sentiment classes and cannot represent neutral or mixed sentiment directly.
• It was trained on English movie reviews and is not a general-purpose emotion or opinion classifier.
• Inputs longer than 256 tokens are truncated, so important conclusions near the end of a long review may be omitted.
• Sarcasm, conflicting opinions, plot-heavy language, and annotation noise can cause errors.
• Confidence scores are model probability estimates, not guarantees that a prediction is correct.
• The model may inherit biases present in its pretrained model and IMDb data.

Source code

Training, evaluation, error-analysis, Slurm, command-line inference, and Gradio code are available in the Sana1025/distilbert-imdb-sentiment-classifier GitHub repository.