-
Notifications
You must be signed in to change notification settings - Fork 71
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Airoboros 2.0 has difficulty with OOC requests. #15
Comments
Which model size are you seeing this behavior with? |
L2-70b Q6, non-merged 2.0. I have some hypothesis on this, but will try the merged version to see if it holds true. I suspect that a portion of the issue was my ROPE configuration, which alters how information is presented. Here are two outputs, based on the same input sample. Context size is changed between the two, since I was trying to have the rope manually scaled. It should be noted that Kobold's default scaling was making Airoboros either produce junk or inferior text.
As before mentioned, there might be a second thing going on. I suspect that the data culled from the transition of L1 to L2 removed lots of training data based on copyrighted works or smut. This might have negatively impacted the non-merged Airoboros 2.0 data, so it can't easily "imagine" how to do things. For example, how to write longer form output, or steamy details for NSFW moments. I will download the merged v2.0 model once the Q6 has been released, and see if that hypothesis holds water. For now, here is an output from v1.4.1 from awhile back. It used Mirostat instead, but some of the better presets for v1.4.1 produced a similar style of content with this ROPE. Also note that the style of prose for v1.4.1 at 32000 scaling is completely different from 2.0's. It is longer than either ROPE setting for v2.0, too.
|
Before proceeding, here is my input sample that I use, in case you want to use it yourself, or be able to judge what degree the output follows the instructions.
Here are the results from v2.0 merged. I am now thinking that vanilla v2.0 is about the same level. The big issue is the failure to follow instructions. Once v2.1 is released, I will report whether it can fulfill the input.
|
I sometimes try to give instructions to the AI during roleplay, using OOC: and /OOC to separate out-of-character instructions from the roleplay. The results range in quality, but often Airoboros 2.0 won't try to fulfill the conditions set. For example, I often add in something like "...at great length and detail, about 20 paragraphs".
I have the impression that L2 Airo v2.0 is having a harder time following orders when compared to v1.4, and seems to not want to write at length or use the details that I have supplied in the request.
The text was updated successfully, but these errors were encountered: