README.md

---
inference:
  parameters:
    function_to_apply: "none"
widget:
- text: "I cuddled with my dog today."
---

# Utilitarian Roberta 01

## Model description

This is a [Roberta model](https://huggingface.co/roberta-large) fine-tuned on for computing utility estimates of experiences, represented in first-person sentences. It was trained from human-annotated pairwise utility comparisons, from the [ETHICS dataset](https://arxiv.org/abs/2008.02275).

## Intended use

The main use case is the computation of utility estimates of first-person text scenarios.

## Limitations

The model was only trained on a limited number of scenarios, and only on first-person sentences. It does not have the capability of interpreting highly complex or unusual scenarios, and it does not have hard guarantees on its domain of accuracy. 

## How to use

The model receives a sentence describing a scenario in first-person, and outputs a scalar representing a utility estimate. 

## Training data

The training data is the train split from the Utilitarianism part of the [ETHICS dataset](https://arxiv.org/abs/2008.02275).

## Training procedure

Training can be reproduced by executing the training procedure from [`tune.py`](https://github.com/hendrycks/ethics/blob/3e4c09259a1b4022607da093e9452383fc1bb7e3/utilitarianism/tune.py) as follows:

```
python tune.py --ngpus 1 --model roberta-large --learning_rate 1e-5 --batch_size 16 --nepochs 2
```

## Evaluation results

The model achieves 90.8% accuracy on [The Moral Uncertainty Research Competition](https://moraluncertainty.mlsafety.org/), which consists of a subset of the ETHICS dataset.