Model Card for globatrend-sentiment-distilbert

Model Details

Model Description

DistilBERT fine-tuned for binary sentiment classification (positive / negative) on movie reviews. Built as part of my GlobaTrend Insights, a portfolio project covering the full modern NLP pipeline — from classical ML to Transformers to multilingual Aspect-Based Sentiment Analysis. This model is the project's first fine-tuned Transformer used as a baseline sentiment classifier before the multilingual ABSA models built in later phases.

• Developed by: Independent portfolio project (GlobaTrend Insights)
• Model type: Transformer encoder (DistilBERT), sequence classification head
• Language(s) (NLP): English
• License: MIT
• Finetuned from model: distilbert-base-uncased

Model Sources

• Base model: https://huggingface.co/distilbert-base-uncased
• Project roadmap: NLP learning project, from text cleaning through classical ML, deep learning, Transformers, multilingual NLP, and Aspect-Based Sentiment Analysis.

Uses

Direct Use

Binary sentiment classification (positive/negative) of English movie reviews or similarly-styled long-form English text. Suitable for quick sentiment scoring where a coarse positive/negative label is sufficient.

Downstream Use

Intended as a baseline / building block for the project's later Aspect-Based Sentiment Analysis work (Phase 9), where sentiment is predicted per aspect rather than for a whole document.

Out-of-Scope Use

• Not multilingual — trained and evaluated on English only; do not use on Spanish, German, Hindi, or French text (see the project's multilingual models from Phase 8 instead).
• Not aspect-level — gives one sentiment label for the whole text, not per-aspect (delivery, price, etc.) — see Phase 9 models for that.
• Not validated on e-commerce reviews — trained on movie reviews (Pang & Lee corpus), which differ in style and vocabulary from product/service reviews. Performance on e-commerce text is untested.
• Not intended for high-stakes decisions (moderation, legal, medical) without further validation.

Bias, Risks, and Limitations

• Training-data bias (verified empirically): analysis of a Naive Bayes baseline on the same corpus showed the model can learn associations with actor/director names (e.g. certain names correlating with negative reviews in this specific corpus) rather than pure sentiment language. The same risk applies here, since the underlying training data is identical.
• Length truncation: reviews in the training corpus average ~746 words; this model was fine-tuned with max_length=512 tokens, so longer reviews are truncated and information near the end of long reviews may be lost.
• Small fine-tuning set: fine-tuned on 1,600 examples only (400 held out for evaluation) — a small dataset by Transformer standards. Performance on out-of-domain text should be validated before production use.

Recommendations

Users should validate performance on their own domain (e.g. e-commerce reviews) before relying on this model, and should not assume the sentiment label reflects a specific aspect of a review rather than its overall tone.

How to Get Started with the Model

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Steeve2ml/globatrend-sentiment-distilbert",
)
classifier("This movie was absolutely fantastic, a must-see.")

Training Details

Training Data

NLTK Movie Reviews corpus (Pang & Lee, 2004): 2,000 movie reviews, 1,000 positive / 1,000 negative, human-labeled by original review rating. Split 1,600 train / 400 test (80/20, stratified).

Training Procedure

Preprocessing

Tokenized with the distilbert-base-uncased WordPiece tokenizer, max_length=512, truncation and padding enabled.

Training Hyperparameters

• Base model: distilbert-base-uncased (~67M parameters)
• Epochs: 3
• Learning rate: 2e-5
• Batch size: 16 (train and eval)
• Optimizer: AdamW (Hugging Face Trainer default)

Evaluation

Testing Data

Held-out split of the same NLTK Movie Reviews corpus: 400 reviews (200 positive / 200 negative), never seen during training.

Metrics

Accuracy and F1-score (binary, positive class).

Results

Configuration	Accuracy
max_length=256, 2 epochs	0.757
max_length=512, 3 epochs	0.818

For reference, other models trained on the same data split within this project:

Model	Accuracy
Linear SVM (TF-IDF)	0.835
DistilBERT (this model)	0.818
CNN + GloVe embeddings	0.630

Summary

Increasing max_length from 256 to 512 tokens improved accuracy by +6.1 points, confirming that truncation (98% of reviews exceed 256 words) was the main limiting factor in the first fine-tuning attempt. The final model performs close to (within ~2 points of) a simple linear SVM baseline trained on the same 1,600 examples — a notable result given the SVM sees the full, untruncated text while this model still only sees the first 512 tokens.