metadata
license: apache-2.0
datasets:
- OpenAssistant/oasst1
language:
- en
🐋 Humback
The proposed Humback is a novel framework that can augment the instruction data for supervised fine-tuning with high quality.
This is a backward model $M_{yx}$ for Humback reproduction.
This model is trained on the seed data in a reversed order (generate instruction given response).
The seed data is a sampled dataset from oasst1.
You may find more details and usage examples in Spico197/Humback .
📜 Reference
@misc{li2023selfalignment,
title={Self-Alignment with Instruction Backtranslation},
author={Xian Li and Ping Yu and Chunting Zhou and Timo Schick and Luke Zettlemoyer and Omer Levy and Jason Weston and Mike Lewis},
year={2023},
eprint={2308.06259},
archivePrefix={arXiv},
primaryClass={cs.CL}
}