Running result

About

This demonstration results from based on the following settings:

1. Pre-trained model: Vicuna 7B in 8 Bit

2. Finetuning adapter model: LoRa

3. Participants: 4 (2 Proposers, 2 Voters)

4. Total communication round: 4

5. Participants' local epoch: 1

6. Stake requirement: 10

7. Total Reward pool: 1000

Training logs

Here are the training logs on the proposer and voter's local PC

Training logs on one proposer (first round role) side

Training logs on one voter (first round role) side

Finetuned model demo

We compared the fine-tuned Vicuna model with the original pre-trained Vicuna model. The results can be found as follows:

Note: Left screen is FLockLLM (Finetuned), right screen is Vicuna (Original)

The self-introduction be finetuned

The finetuned model shows a preference for answering domain questions

The finetuned LLM could answer most recent events/news

The finetuned model still has the similiar performance even better for answering general QA.