harshaojha/bert-imdb-finetuned

BERT — IMDB Sentiment Classifier

A fine-tuned bert-base-uncased model that classifies English text as Positive or Negative sentiment. Trained on the IMDB movie reviews dataset.

🎛️ Live demo: huggingface.co/spaces/harshaojha/bert-imdb-demo

Model details

• Base model: bert-base-uncased (110M parameters)
• Task: Binary text classification (sentiment)
• Labels: 0 = Negative, 1 = Positive
• Language: English
• Max input length: 256 tokens

Intended use

Educational / demonstration — learning how to fine-tune a transformer and wrap it in a web UI. Works well on movie-review-style text; less reliable on other domains.

Training data

A balanced subset of the IMDB reviews dataset:

• Train: 10,000 reviews (shuffled, seed=42)
• Validation: 2,000 reviews (shuffled, seed=42)

Training procedure

Hyperparameter	Value
Optimizer	AdamW
Learning rate	2e-5
Batch size (train / eval)	16 / 32
Epochs	3
Weight decay	0.01
Warmup ratio	0.1
Max sequence length	256
Best-checkpoint metric	accuracy
Hardware	NVIDIA T4 GPU (Google Colab)

Evaluation results

Epoch	Validation Accuracy
1	89.80%
2	91.45% ← best
3	91.10%

How to use

from transformers import pipeline

clf = pipeline("text-classification", model="harshaojha/bert-imdb-finetuned")
ID2LABEL = {"LABEL_0": "Negative", "LABEL_1": "Positive"}
result = clf("An absolute masterpiece.")[0]
print(ID2LABEL[result["label"]], round(result["score"], 3))
# Positive 0.993

Or load the tokenizer and model directly:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("harshaojha/bert-imdb-finetuned")
model = AutoModelForSequenceClassification.from_pretrained("harshaojha/bert-imdb-finetuned")
inputs = tokenizer("Painfully boring.", return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)[0]
print({"Negative": probs[0].item(), "Positive": probs[1].item()})

Limitations

• Trained on movie reviews only — accuracy drops on other domains (tweets, product reviews, news).
• Inherits biases present in IMDB and in the base bert-base-uncased model.
• Truncates inputs longer than 256 tokens.
• Trained on 10,000 of the 25,000 available training samples; using the full set would likely push accuracy higher.

Author

Trained by @harshaojha as part of a learning project on transformer fine-tuning.