VijayKanupuri/bert-sst2

BERT Fine-tuned on SST-2

Model Description

This model is a BERT-base (bert-base-uncased) model fine-tuned on the SST-2 (Stanford Sentiment Treebank) dataset for binary sentiment classification.

The model predicts whether a given English sentence expresses positive or negative sentiment.

This model was fine-tuned as part of a learning-focused, production-style workflow:

• training on GPU (Google Colab)
• exporting model artifacts
• running inference locally

---

Intended Use

Supported Use Cases

• Sentiment analysis of short English sentences
• Educational and learning purposes
• Benchmarking text classification pipelines
• Prototyping NLP applications

Out-of-Scope Uses

• Long documents without chunking
• Multilingual sentiment analysis
• High-stakes decision making (legal, medical, financial)

---

Training Data

• Dataset: SST-2 (Stanford Sentiment Treebank)
• Source: GLUE benchmark
• Task: Binary sentiment classification
• Labels:
  • 0 → Negative
  • 1 → Positive

SST-2 consists of movie review sentences annotated for sentiment polarity.

---

Training Details

• Base Model: bert-base-uncased
• Architecture: Encoder-only Transformer (BERT)
• Number of Labels: 2
• Loss Function: Cross-Entropy Loss
• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 32 (GPU)
• Epochs: 2
• Precision: Mixed precision (fp16)
• Training Hardware: NVIDIA GPU (Google Colab)

---

Evaluation

The model was evaluated on the SST-2 validation set.

Expected performance is consistent with standard BERT fine-tuning on SST-2:

• Accuracy: ~90% (depending on training seed and subset)

This model is intended primarily for learning and demonstration rather than leaderboard optimization.

---

How to Use

Inference Example

from transformers import BertTokenizer, BertForSequenceClassification
import torch

tokenizer = BertTokenizer.from_pretrained("VijayKanupuri/bert-sst2")
model = BertForSequenceClassification.from_pretrained("VijayKanupuri/bert-sst2")

text = "I really loved this movie"

inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)

label = torch.argmax(probs, dim=-1).item()
confidence = probs[0][label].item()

print(label, confidence)