A state-of-the-art sentiment analysis model achieving 89.2% accuracy on IMDB and 91.5% on Stanford SST-2
RoBERTa-Sentimentic is a fine-tuned RoBERTa model specifically optimized for sentiment analysis across multiple domains. Trained on 50,000+ samples from IMDB movie reviews and Stanford Sentiment Treebank, it demonstrates exceptional performance in binary sentiment classification with robust cross-domain transfer capabilities.
from transformers import pipeline
# Load the model
classifier = pipeline("sentiment-analysis", model="abhilash88/roberta-sentimentic")
# Single prediction
result = classifier("This movie is absolutely fantastic!")
print(result)
# [{'label': 'POSITIVE', 'score': 0.998}]
# Batch predictions
texts = [
"Amazing cinematography and outstanding performances!",
"Boring plot with terrible acting.",
"A decent movie, nothing extraordinary."
]
results = classifier(texts)
for text, result in zip(texts, results):
print(f"Text: {text}")
print(f"Sentiment: {result['label']} (confidence: {result['score']:.3f})")
| Dataset | Pre-trained RoBERTa | RoBERTa-Sentimentic | Improvement |
|---|---|---|---|
| IMDB Movie Reviews | 49.5% | 89.2% | +39.7% |
| Stanford SST-2 | 49.1% | 91.5% | +42.4% |
| Cross-domain (IMDB→SST) | 49.1% | 87.7% | +38.6% |
Predicted
Actual Negative Positive
Negative 2789 336
Positive 341 2784
Precision: 89.2% | Recall: 89.1% | F1-Score: 89.1%
Predicted
Actual Negative Positive
Negative 412 16
Positive 58 386
Precision: 91.5% | Recall: 91.4% | F1-Score: 91.5%
| Metric | Pre-trained | Fine-tuned | Improvement |
|---|---|---|---|
| IMDB Accuracy | 49.5% | 89.2% | 🔥 +80.2% relative |
| SST-2 Accuracy | 49.1% | 91.5% | 🔥 +86.4% relative |
| Average Confidence | 0.51 | 0.94 | +84.3% |
| Error Rate | 50.7% | 9.6% | -81.1% |
Model: roberta-base
Fine-tuning Strategy: Domain-specific + Cross-domain validation
Training Samples: 50,000+ (IMDB: 25k, SST-2: 25k)
Hyperparameters:
Learning Rate: 2e-5
Batch Size: 16
Epochs: 3
Weight Decay: 0.01
Warmup Steps: 200
Max Length: 256 tokens
Optimization:
Optimizer: AdamW
Scheduler: Linear with warmup
Loss Function: CrossEntropyLoss (with class weights for SST-2)
Hardware: NVIDIA GPU (Google Colab)
Training Time: ~25 minutes total
| Model | IMDB Accuracy | SST-2 Accuracy | Parameters |
|---|---|---|---|
| RoBERTa-Sentimentic | 89.2% | 91.5% | 125M |
| RoBERTa-base (pre-trained) | 49.5% | 49.1% | 125M |
| BERT-base-uncased | ~87.0% | ~88.0% | 110M |
| DistilBERT-base | ~85.5% | ~86.2% | 67M |
pip install transformers torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline
# Method 1: Using pipeline (recommended)
classifier = pipeline("sentiment-analysis", model="abhilash88/roberta-sentimentic")
result = classifier("Your text here")
# Method 2: Direct model usage
tokenizer = AutoTokenizer.from_pretrained("abhilash88/roberta-sentimentic")
model = AutoModelForSequenceClassification.from_pretrained("abhilash88/roberta-sentimentic")
inputs = tokenizer("Your text here", return_tensors="pt", truncation=True)
outputs = model(**inputs)
predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
import torch
from transformers import pipeline
# Load model with specific device
device = 0 if torch.cuda.is_available() else -1
classifier = pipeline(
"sentiment-analysis",
model="abhilash88/roberta-sentimentic",
device=device
)
# Batch processing for efficiency
texts = ["Text 1", "Text 2", "Text 3", ...]
results = classifier(texts, batch_size=32)
# Get raw confidence scores
for text, result in zip(texts, results):
label = result['label']
confidence = result['score']
print(f"Text: {text}")
print(f"Sentiment: {label} (confidence: {confidence:.3f})")
precision recall f1-score support
NEGATIVE 0.89 0.89 0.89 3125
POSITIVE 0.89 0.89 0.89 3125
accuracy 0.89 6250
macro avg 0.89 0.89 0.89 6250
weighted avg 0.89 0.89 0.89 6250
precision recall f1-score support
NEGATIVE 0.92 0.96 0.94 428
POSITIVE 0.96 0.87 0.91 444
accuracy 0.92 872
macro avg 0.94 0.91 0.92 872
weighted avg 0.94 0.92 0.92 872
# IMDB Dataset Processing
imdb_train: 25,000 samples (balanced: 50% positive, 50% negative)
imdb_test: 6,250 samples
# Stanford SST-2 Processing
sst_train: 67,349 samples → sampled 25,000 (balanced)
sst_validation: 872 samples (used for evaluation)
# Label Standardization
IMDB: {0: "NEGATIVE", 1: "POSITIVE"} ✓
SST-2: {-1: "NEGATIVE", 1: "POSITIVE"} → {0: "NEGATIVE", 1: "POSITIVE"} ✓
If you use this model in your research, please cite:
@misc{roberta-sentimentic,
title={RoBERTa-Sentimentic: Fine-tuned Sentiment Analysis with Cross-Domain Transfer},
author={Abhilash},
year={2025},
publisher={Hugging Face},
journal={Hugging Face Model Hub},
howpublished={\url{https://huggingface.co/abhilash88/roberta-sentimentic}}
}
This model is released under the Apache 2.0 License. See LICENSE for details.
🌟 If this model helped your project, please give it a ⭐ star! 🌟
checkpoint-400/optimizer.pt 997 MBcheckpoint-800/optimizer.pt 997 MBcheckpoint-939/optimizer.pt 997 MBcheckpoint-400/model.safetensors 499 MBcheckpoint-800/model.safetensors 499 MBcheckpoint-939/model.safetensors 499 MBmodel.safetensors 499 MBcheckpoint-400/rng_state.pth 14.2 kBcheckpoint-800/rng_state.pth 14.2 kBcheckpoint-939/rng_state.pth 14.2 kBcheckpoint-400/training_args.bin 5.4 kBcheckpoint-800/training_args.bin 5.4 kBcheckpoint-939/training_args.bin 5.4 kBtraining_args.bin 5.4 kBcheckpoint-400/scheduler.pt 1.1 kBcheckpoint-800/scheduler.pt 1.1 kBcheckpoint-939/scheduler.pt 1.1 kBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo abhilash88/roberta-sentimentic ./model-folderFacebookAI/roberta-base