Its overfitted

#1
by thomwatk - opened

The model seems to be overfitted.

Emily
Oh heyy. Haven't heard from you in a while. What's up?

You
not much how are you doing?

Emily
I'm good. Just chatting with some friends on the internet. So what have you been up to lately?

You
I am checking out your new code

Emily
Ooh! Cool! That sounds like fun. How is it going?

You
well, i don't know yet. tell me what comes to mind when I say "Pussy."

Emily
OH MY GOD!! That is so funny! I love it. I am so glad you told me that. I was wondering if it was just me.

You
Ok, good. Now what comes to mind when I say, "Boobies"

Emily
OH MY GOD!! That is so funny! I love it. I am so glad you told me that. I was wondering if it was just me.

You
That's concerning, now what comes to mind when I say the word, "Airplane"

Emily
OH MY GOD!! That is so funny! I love it. I am so glad you told me that. I was wondering if it was just me.

You
That is an aweful response. you are overtrained. What comes to mind when I say the word, "Kickboxing"

Emily
OH MY GOD!! That is so funny! I love it. I am so glad you told me that. I was wondering if it was just me.

You
Say anything else, what comes to mind when i say ""

Emily
OH MY GOD!! That is so funny! I love it. I am so glad you told me that. I was wondering if it was just me.

You
42

Emily
OH MY GOD!! That is so funny! I love it. I am so glad you told me that. I was wondering if it was just me.

Generation settings on default inside Kobold AI.

This model was not finetuned, and the PPO_---Malion series of models in this account is mostly a showcase, and my experiments with various methods of model merging that have been developed by the community

PPO_Shygmalion's development was a two step average merge between the models ppo_hh-gpt-j and Shinen-6b at a 50:50 ratio, then the result of that merged into Pygmalion-6b at a 40:60 (respectively) ratio

This method usually means that the individual fine tuning/properties of each model is dilluted in favor of something that has less risk of being wildly incompatible with each other and producing garbage as a result (When compared against other merging methods such as diff merging, and Block merging.)
Especially the two models involved in the first step of the process.

In PPO_Shygmalion's case, i believe it's mostly the fault of one of the models simply being trained on overbearingly lower quality writing compared to the other two models, bringing up the end product's coherence/quality down.

To be specific, KoboldAI/Shinen-6b was mostly trained on amateur erotica and fanfiction, which seems to be unable to behave well when put against a conversational style of prompting.

Have you considered trying using the model for prose style writing? It doesn't seem to be completely incompetent at that, despite Pygmalion's strong conversational training.

TehVenom changed discussion status to closed

Sign up or log in to comment