APACHE-2.0 License
"If there was a robot that could laugh, cry and smile. Does it have a soul?"
Join our QQ Group.
In the past decades, romance between humans and androids has become a popular theme in numerous literature and visual novels. Readers may imagine that there will be artificial intelligence (AI) capable of experiencing authentic emotions in the near future. Recently, with the rapid development of pre-trained language models (PLMs), ChatGPT and its analogues demonstrate stunning performance in communication with humans and especially in alignment with human preferences. However, the primary objective of these robots is to follow the human instructions so as to serve as an intelligent assistant. These robots lack characteristics and hence humans can hardly empathize with them. Although it is possible to instruct AI to perform a role-play by designing prompts containing the description of a virtual character, there are several drawbacks of this approach:
Despite the aforementioned limitations, AI has the potential of leveraging pretrained knowledge to imitate any virtual character. The effectiveness of prompts is questionable since the knowledge contained in AI remains unaffected. On the contrary, we conjecture that fine-tuning the PLMs under the objective of aligning their behaviors with virtual characters using the adequate corpus can achieve satisfactory results. Therefore, in this project, we aim to fine-tune the PLMs with state-of-the-art training schemes to generate AIs that behaves like arbitrary virtual characters. This will enable the realization of the dream presented in the visual novels.
Firstly, we will utilize the script from the visual novels to generate datasets for fine-tuning with the following steps:
The dataset includes two types:
name
, age
and hobbies
.Now we are handling the following visual novels:
We are based on the pre-trained large language models that trained on massive Chinese corpus. Inspired by ChatGPT, the paradigm of fine-tuning we adopted is supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF). Our fine-tuning largely depends on the LLaMA-Efficient-Tuning and ChatGLM-Efficient-Tuning framework. Please refer to these repos for more details.
The generated AI can equip with the following multimodal features:
We suppose this project may have promising potential and broad impacts because the following advantages:
This repository is licensed under the Apache-2.0 License.
This repo is based on LLaMA-Efficient-Tuning and ChatGLM-Efficient-Tuning.