Some errors at runtime

[renhl717445]

When I run sh demo_bisr.sh in the pippline directory, the following errors are reported: cannot import name 'gaussian_moe_range' from 'sfl.utils.args' (/home/LLM/SplitLLM/sfl/utils/args.py), and the modules gaussian_moe_range, lora_path, dc_moe_range are not retrieved in the args.py file; in addition, the error No module named 'sfl.model.noise.fdp' is reported. Is there no fdp.py file?

[StupidTrees]

We apologize for the current issues! This repository is still under construction🥺😢
In future updates, we will review the issues, organize the code, and add more detailed guide documentation.

[renhl717445]

In sim_with_attack.py, it seems that the pre-trained model is directly reversed for input recovery. Is there no reverse step after fine-tuning the model? Are all the fine-tuning processes in the paper done for frozen server-side models? simulator.pre_ft is federated training after fine-tuning, and does not return information about the model. Where should I modify it?

[StupidTrees]

By default, the simulator performs a data reconstruction attack during the fine-tuning process.
For example, after setting attacker_freq to 500, the simulator will invoke Attackers to launch an attack after every 500 steps of fine-tuning.

The pre_ft option provides a way to conduct several rounds of fine-tuning on a different dataset before the simulation experiment, thereby shifting the weights of the LLM.