DistilBERT-base-uncased fine-tuned for binary sentiment classification (positive/negative movie reviews). A genuine transformer fine-tune performed locally on a 1,500-review IMDB subsample.

Model Characteristics

Base model: distilbert-base-uncased (Hugging Face)
Task: Binary sequence classification
Eval loss: 0.0005 (after early stopping on epoch 2)
Training: 2 epochs, lr 2e-5, batch 8, warmup 50 steps

Data Overview

Dataset: stanfordnlp/imdb (subsampled to 1,500 train / 300 test for CPU feasibility)
Text truncated/padded to 128 tokens

Evaluation Results

Eval loss: 0.000518
Early stopping triggered at epoch 2 of 3

Files

model.safetensors — fine-tuned weights
config.json — model config (2 labels)
tokenizer.json, tokenizer_config.json — tokenizer
eval_metrics.txt — final evaluation metrics

Usage

python

from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained("bolewara/imdb-sentiment-distilbert")
tokenizer = AutoTokenizer.from_pretrained("bolewara/imdb-sentiment-distilbert")

Writeup

A full case study is in the accompanying writeup (also published on Kaggle): Fine-Tuning DistilBERT for Movie Review Sentiment Analysis.

Provenance

Fine-tuned by Anuj Bolewar (anujbolewar on Kaggle) from distilbert-base-uncased on stanfordnlp/imdb.