Sanjeev2501/nyxar-roberta-sentiment

NYXAR RoBERTa Sentiment Model

Model Description

A transformer-based sentiment classification model developed for the NYXAR AI Intelligence & Observability Platform. The model leverages a fine-tuned RoBERTa architecture optimized for customer feedback intelligence and predicts sentiment across three classes: Positive, Neutral, and Negative.

Framework: Hugging Face Transformers, ONNX Runtime, RoBERTa Language: English License: MIT Base Model: roberta-base

---

Training Data

The model was trained on SetFit/amazon_reviews_multi_en, an English Amazon reviews dataset commonly used for sentiment classification tasks.

---

Intended Use

This model is designed for customer feedback analysis, product review monitoring, support ticket intelligence, sentiment trend analysis, and enterprise AI intelligence workflows.

---

Limitations

The model may struggle with sarcasm, irony, ambiguous sentiment expressions, domain-specific terminology not represented in the training data, and highly subjective reviews. Predictions should be used as supporting signals rather than business-critical decisions.

---

Performance

Metric	Score
Accuracy	77.56%
Precision	77.99%
Recall	77.56%
F1 Score	77.76%

---

Usage

Assumes tokenizer and model_quantized.onnx are in ./onnx/

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer
from scipy.special import softmax

model_path = "./onnx"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path)

# Load ONNX model
session = ort.InferenceSession(
    f"{model_path}/model_quantized.onnx"
)

text = "The product exceeded expectations."

# Tokenize
inputs = tokenizer(
    text,
    return_tensors="np",
    truncation=True,
    padding=True
)

# Convert tokenizer outputs to ONNX inputs
ort_inputs = {
    k: v.astype(np.int64)
    for k, v in inputs.items()
}

# Run inference
outputs = session.run(None, ort_inputs)

# Get logits
logits = outputs[0]

# Compute probabilities
probs = softmax(logits, axis=-1)

prediction = int(np.argmax(probs, axis=-1)[0])
confidence = float(np.max(probs))

print(prediction, confidence)

---