| --- |
| language: multilingual |
| license: apache-2.0 |
| tags: |
| - sentiment-analysis |
| - text-classification |
| - xlm-roberta |
| - dual-head |
| --- |
| |
| # Sentiment Classifier |
|
|
| ## Model Description |
|
|
| This is a dual-head sentiment classifier built on top of XLM-RoBERTa. The model performs two tasks simultaneously: |
|
|
| 1. **Sentiment Classification:** Predicts sentiment labels (positive, neutral, negative) |
| 2. **Sentiment Score Regression:** Predicts a continuous sentiment score in the range [0, 1] |
|
|
| The model uses a weighted loss function combining cross-entropy (70%) for classification and MSE (30%) for regression, |
| allowing it to capture both discrete sentiment categories and fine-grained sentiment intensity. |
|
|
| ## Model Architecture |
|
|
| - **Base Model:** xlm-roberta-base |
| - **Task:** text-classification |
| - **Number of Labels:** 3 |
| - **Labels:** negative, neutral, positive |
|
|
| ## Training Configuration |
|
|
| - **Epochs:** 10 |
| - **Batch Size:** 128 |
| - **Learning Rate:** 2e-05 |
| - **Warmup Ratio:** 0.1 |
| - **Weight Decay:** 0.01 |
| - **Max Seq Length:** 256 |
| - **Mixed Precision:** FP16=False, BF16=True |
|
|
| ## Performance Metrics |
|
|
| - **Loss:** 0.6947 |
| - **Accuracy:** 0.4901 |
| - **Precision:** 0.2402 |
| - **Recall:** 0.4901 |
| - **F1:** 0.3224 |
| - **F1 Macro:** 0.3289 |
| - **F1 Negative:** 0.0000 |
| - **Precision Negative:** 0.0000 |
| - **Recall Negative:** 0.0000 |
| - **Support Negative:** 900 |
| - **F1 Neutral:** 0.6578 |
| - **Precision Neutral:** 0.4901 |
| - **Recall Neutral:** 1.0000 |
| - **Support Neutral:** 865 |
| - **Runtime:** 0.7012 |
| - **Samples Per Second:** 2517.1350 |
| - **Steps Per Second:** 9.9830 |
|
|
| ## Model Outputs |
|
|
| The model returns two outputs: |
|
|
| - **Logits:** Classification logits for sentiment labels [batch_size, 3] |
| - **Score Predictions:** Continuous sentiment scores [batch_size] |
|
|
| Both outputs are computed from the same shared representation (CLS token) of the input text. |
|
|
| ## Intended Use |
|
|
| This model is intended for sentiment analysis tasks on multilingual text, particularly in scenarios where both |
| categorical sentiment (positive/neutral/negative) and sentiment intensity are important. |
|
|
| **Typical use cases:** |
| - Product review analysis |
| - Social media sentiment monitoring |
| - Customer feedback classification |
|
|
| ## Usage |
|
|
| ```python |
| from transformers import AutoTokenizer |
| from src.models.sentiment_classifier import SentimentClassifier |
| |
| # Load model and tokenizer |
| model = SentimentClassifier.from_pretrained("YOUR_USERNAME/sentiment-classifier") |
| tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base") |
| |
| # Prepare input |
| text = "Your input text here" |
| inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True) |
| |
| # Make prediction |
| outputs = model(**inputs) |
| predictions = outputs["logits"].argmax(dim=-1) |
| ``` |
|
|
| ## Citation |
|
|
| If you use this model, please cite: |
|
|
| ```bibtex |
| @misc{sentiment_classifier, |
| title={Sentiment Classifier}, |
| author={{Your Name}}, |
| year={2025}, |
| publisher={Hugging Face}, |
| howpublished={{\url{{https://huggingface.co/YOUR_USERNAME/{model_name.lower().replace(' ', '-')}}}}} |
| } |
| ``` |
|
|
| --- |
|
|
| *This model card was automatically generated with [Claude Code](https://claude.com/claude-code)* |
