YueranCao2001/distilbert-imdb-sentiment-v3

DistilBERT IMDb Sentiment Classifier (V3)

Model Description

This model is a fine-tuned version of distilbert/distilbert-base-uncased for binary sentiment classification of English movie reviews.

It was developed as an independent Hugging Face learning project, covering dataset preparation, tokenization, fine-tuning, checkpoint selection, evaluation, and model publishing.

The model predicts two sentiment labels:

• NEGATIVE (0)
• POSITIVE (1)

Training Dataset

Dataset: IMDb Movie Reviews

The training and validation samples were drawn from the official IMDb training split.

Split	Samples
Training	5,000
Validation	1,000
Independent test subset	200

The training and validation datasets were balanced across positive and negative classes.

The test subset was sampled from the official IMDb test split, with 100 examples per class.

Training Configuration

Parameter	Value
Base model	DistilBERT Base Uncased
Epochs	3
Training batch size	16
Evaluation batch size	16
Learning rate	2e-5
Weight decay	0.01
Maximum sequence length	128
Learning rate scheduler	Linear
Warmup steps	94
Random seed	42
Model selection metric	Validation Macro F1
Best checkpoint	Step 939
Training device	CPU

The model was trained using Hugging Face Transformers and PyTorch.

Evaluation Results

Validation Results

Metric	Value
Accuracy	0.8500
Macro F1	0.84998
Loss	0.42117

IMDb Test Subset (200 Samples)

Metric	Value
Accuracy	0.86
Macro F1	0.86
Negative Recall	0.84
Positive Recall	0.88

Confusion matrix:

[[84, 16],
 [12, 88]]

These results are based on a small, balanced subset of the official IMDb test split. They should not be interpreted as performance on the complete IMDb benchmark.

The same test subset was also used for comparison with earlier experimental versions, so these results are exploratory rather than a fully untouched final benchmark.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="YOUR_USERNAME/distilbert-imdb-sentiment-v3"
)

result = classifier(
    "This movie was absolutely fantastic!"
)

print(result)

Replace YOUR_USERNAME with the actual Hugging Face account name after publishing.

Limitations

• Designed for English movie-review sentiment classification.
• Supports only positive and negative labels.
• Does not explicitly recognize neutral sentiment.
• Input text is truncated to 128 tokens.
• Performance may degrade on other domains.
• Evaluation was conducted on a limited test subset.
• Intended primarily for learning and demonstration.

License

Apache-2.0, following the base model's license.

Project Scope

This is an independent Hugging Face learning project and does not depend on an external GitHub repository.