README.md

---
language: en
license: mit
tags:
  - sentiment-analysis
  - text-classification
  - modernbert
library_name: sentimentizer
task: text-classification
---
# Sentimentizer MODERNBERT Sentiment Model
## Description

A ModernBERT-base model fine-tuned for 3-class sentiment classification (negative, neutral, positive). The backbone uses mean pooling over non-padding tokens followed by a two-layer classifier head with GELU activation. Supports 8-bit AdamW optimization and layer-wise unfreezing.

## Training Data

Trained on the [Yelp Open Dataset](https://www.yelp.com/dataset) reviews with 3-class labels (negative/neutral/positive). The ModernBERT tokenizer handles subword tokenization natively—no custom dictionary is needed.

## Usage

```python
from sentimentizer.hf import download_weights
from sentimentizer.config import weights_path_for

# Download classifier head + backbone from Hugging Face Hub
weights_path = weights_path_for("modernbert")
download_weights(
    "modernbert",
    weights_path,
    repo_id="ryeyoo/sentimentizer-modernbert",
)

# Load and run inference
from sentimentizer.models.modernbert import new_modernbert_model
from sentimentizer.predictor import SentimentPredictor

model = new_modernbert_model()
predictor = SentimentPredictor(model)

result = predictor.predict('amazing food great service')
print(result)
# e.g. {'label': 'positive', 'score': 0.83, ...}
```

## Files

- `modernbert_weights.pth` — Classifier head weights and config metadata
- `backbone/` — Hugging Face transformer backbone (safetensors + config)
- `backbone/tokenizer.json` — Full tokenizer data
- `backbone/tokenizer_config.json` — Tokenizer configuration