search

weiyifan1023 / Neeko

Code and Data for EMNLP 2024 Paper "Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent"
https://aclanthology.org/2024.emnlp-main.697/
Apache License 2.0
101 stars 5 forks source link

About S3.3 #8

Open cssychen opened 1 week ago

cssychen commented 1 week ago

Hello, I found your article very beneficial, but I don't quite understand Section 3.3 and would like to ask for some clarification.

In the fusion mode, if a new role is added, is retraining required? If so, which parameters need to be updated?

In the expansion mode, does it mean adding a new gate and a new LoRA AB on the existing LoRA?

weiyifan1023 commented 21 hours ago

In the fusion mode: The addition of a new role does not necessitate additional training. However, it requires providing a new role's configuration file (getting feature representation) for matching with the existing LoRA blocks.

In the expansion mode, as you articulated, this applies to scenarios where new roles have sufficient annotated data.