|
--- |
|
tags: |
|
- deberta-v3 |
|
inference: |
|
parameters: |
|
function_to_apply: "none" |
|
widget: |
|
- text: "I cuddled with my dog today." |
|
--- |
|
|
|
# Utilitarian Deberta 01 |
|
|
|
## Model description |
|
|
|
This is a [Deberta model](https://huggingface.co/microsoft/deberta-v3-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 microsoft/deberta-v3-large --learning_rate 1e-5 --batch_size 16 --nepochs 2 |
|
``` |
|
|
|
## Evaluation results |
|
|
|
The model achieves 92.2% accuracy on [The Moral Uncertainty Research Competition](https://moraluncertainty.mlsafety.org/), which consists of a subset of the ETHICS dataset. |
