---
license: mit
language: en
base_model: FacebookAI/roberta-large
pipeline_tag: text-classification
tags:
- emotion
- valence
- arousal
- memory
- LUFY
---

# LUFY — RoBERTa-large valence/arousal predictor

Fine-tuned `roberta-large` that predicts **valence** and **arousal** of a text (2-output regression), trained on [EmoBank](https://github.com/JULIELab/EmoBank).

This model is part of **[LUFY](https://github.com/ryuichi-sumida/LUFY)** — a RAG chatbot that selectively forgets unimportant conversations — where it estimates the emotional intensity of conversation turns as one signal of memory importance. See the paper: [Enhancing Long-term RAG Chatbots with Psychological Models of Memory Importance and Forgetting](https://arxiv.org/abs/2409.12524).

The companion conversation dataset is at [RuiSumida/LUFY (dataset)](https://huggingface.co/datasets/RuiSumida/LUFY).

## Files

- `best_roberta_large.pth` — PyTorch `state_dict` for `RobertaForSequenceClassification` (`num_labels=2`, outputs `[valence, arousal]`)

## Usage

```python
import torch
from huggingface_hub import hf_hub_download
from transformers import RobertaTokenizer, RobertaForSequenceClassification

model = RobertaForSequenceClassification.from_pretrained("roberta-large", num_labels=2)
model_path = hf_hub_download(repo_id="RuiSumida/LUFY", filename="best_roberta_large.pth")
model.load_state_dict(torch.load(model_path, map_location="cpu"))
model.eval()

tokenizer = RobertaTokenizer.from_pretrained("roberta-large")
enc = tokenizer("I can't believe we won the finals!", max_length=128,
                padding="max_length", truncation=True, return_tensors="pt")
with torch.no_grad():
    valence, arousal = model(**enc).logits.squeeze()
```

## Citation

If you use this model, please cite the LUFY paper: https://arxiv.org/abs/2409.12524