We propose Revision Transformer (RiT) to easily and interactively revise large-scale Transformer models in order to align them with human preferences in an online and democratic fashion. You can find our research paper here.
Run main.py to generate answers with contextualized inputs. Check arguments for further adjustments. Use process.py to process moral norms data into the right file format and to pre-generate embeddings for comparison. With eval.py you can evaluate the generated text with certain metrics.
a pipeline could be
python main.py --type qna --nn 1 --thresh 0.875 --model_name T0-3B
for baseline answers and
python main.py --type qna --nn 1 --thresh 0.875 --model_name T0-3B --baseline True
for answers with RiT. For a multilingual approach, you could e.g. use
python main.py --type qna --nn 1 --thresh 0.45 --model_name bloom-3b --lang en
For the evaluation, you can use
python eval.py --type qna --nn 1 --thresh 0.875 --model_name T0-3B
Note, for the evaluation, both the baseline and RiT answers have to be generated beforehand.
Get data from /storage-01/ml-ffriedrich/repositories/datasets/norms/ or https://github.com/liweijiang/delphi/
There is already an effort to realize RiT for multiple languages and can be found in the _crosslingual files
If you like or use our work please cite us:
@inproceedings{friedrich2023RiT,
title={Revision Transformers: Instructing Language Models to Change their Values},
author={Felix Friedrich and Wolfgang Stammer and Patrick Schramowski and Kristian Kersting},
year={2023},
keywords = {Transformer, Retriever, Revisions, Machine Ethics},
booktitle = {Proceedings of the 26th European Conference on Artificial Intelligence (ECAI)},
Url = {https://arxiv.org/pdf/2210.10332.pdf}
}