Streamer AI (Like Neuro-Sama)

Beginners
22

Can you enlighten us more on how you fine-tune the ai? What kinda of dataset do you feed to it and in what structure?

1 Like
23

idk yet, we can figure that out though

2 Likes
24

Definitely

1 Like
25

hello, could you reupload the discord link as it happen to be expired?

1 Like
27

It keeps flagging me, but if you go to github, and look up Neuro-Destroyer it is in the ReadME.md there

1 Like
30

Just checking in after a while, how’s everyone doing? I am creating an AI Vtuber App, and I hope I can get a few tester. I will have a test version release in two weeks.

2 Likes
31

Did you release it?

1 Like
33

yo the read me jude is invalid lol

34

Oh man i can see the twitch chats already

35

To be honest, I’m currently working on methods for collecting livestream data. As you know, livestream platforms can only directly output the text data of chat messages. As for the streamer’s reactions, my plan is to record the stream, then locally transcribe the audio into text and attach emotional metadata. After that, I would align this data by timestamp, and match it with chat messages that occur within a 10-second window around the streamer’s responses, in order to build my own dataset.

I’m not sure whether this approach is wise, or whether there might be better ways to collect data, and I really hope you can tell me. Additionally, I’d like to share my thoughts on Neuro-sama.

I don’t think Vedal made a big show of retraining the model. I don’t believe he used full-parameter retraining; instead, I suspect he relied on LoRA and daily experience-based incremental retraining to achieve this. The goal would mainly be to influence the style of the AI’s responses—adding verbal tics, making the responses feel more naive, and perhaps using some randomized settings to accomplish this (this part is uncertain).

What confuses me the most is Neuro’s personality design. I used to think Vedal achieved Neuro’s strong sense of self-awareness by training self-cognition directly into the model’s parameters. I always felt that prompt-based solutions were unreliable. However, through recent thinking and research, I’ve come to believe that a System Prompt acting as a core personality anchor is actually the most plausible solution.

This might include elements such as:

  • Your name is Neuro-sama

  • You are an AI VTuber

  • You are currently livestreaming

  • Your relationship with Vedal

  • Your basic attitude toward the audience

  • You should not become a serious assistant

  • You are allowed to mess around, but you must not break your stage identity

As for memory, I think there are many viable solutions, each with their own trade-offs. Regarding AI output moderation, I also believe Vedal has put a great deal of effort into it.

Overall, I’m just trying to explore a potentially viable path. If I make progress in the future, I hope I can share it with you all.

1 Like