shawnnygoh/cs4248-roberta-sentiment

Fine-tuned RoBERTa for Tweet Sentiment (3-class)

Fine-tuned RoBERTa for 3-class sentiment classification (negative, neutral, positive) on the Tweet Sentiment Analysis Dataset (TSAD).

Usage

Pipeline

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="shawnnygoh/cs4248-roberta-sentiment",
)

results = classifier([
    "This is amazing!",
    "Okay",
    "Absolutely terrible experience",
])

for r in results:
    print(f"{r['label']}: {r['score']:.4f}")
# positive: 0.9907
# neutral: 0.9870
# negative: 0.9770

AutoTokenizer + AutoModel

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "shawnnygoh/cs4248-roberta-sentiment"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

texts = [
    "Generational trauma. An example of why forgiveness is so important.",
    "Improved longform posts",
    "This is the worst thing I've ever seen",
]

inputs = tokenizer(texts, padding=True, truncation=True, max_length=128, return_tensors="pt")

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

label_names = ["negative", "neutral", "positive"]
for text, pred, prob in zip(texts, preds, probs):
    print(f"{label_names[pred]}: {prob[pred]:.4f} | {text}")

Training details

Parameter	Value
Base model	FacebookAI/roberta-base
Dataset	TSAD (3-class: negative, neutral, positive)
Train/test split	85% / 15%, stratified
Validation	15% of train, used for early stopping
Epochs	10 (early stopping patience: 2)
Batch size	16
Learning rate	1e-5
Weight decay	0.01
Warmup ratio	0.1
Max sequence length	128
Precision	fp16
Best model selection	macro F1 on validation set

Preprocessing

Minimal cleaning applied before tokenization: URL removal, @mention removal, whitespace normalization, and expressive lengthening collapse (e.g. s000000 → s000).

Label mapping

Label	ID
negative	0
neutral	1
positive	2