SanjiWatsuki

/

Loyal-Macaroni-Maid-7B

like 56

Thanks for testing the model, I'm glad you're having a great time! :)

Thanks for testing the model, I'm glad you're having a great time! :)

He's right, this is literally one of the best models i've tested for roleplay. In my case i'm using The Blokes AWQ version but since this is a 7B model, i'm going to use the full 16 bit version from now on.

The model's memory is very on point, the output quality is also incredible. However from my testing, not sure if it's my character cards or if it's the AWQ or something else, but i did notice the model tries to "Curve" away from the NSFW stuff sometimes and it tries to avoid it. At that point i would have to "Convince" the character to do it. So not sure what that was about, but i'll test it with the full 7B model soon. Other than that, this is so far the best roleplay model i've tested.

i was prob gonna ask if there was going to be a V2 or a 13B version but given the fact most of the models used for this were 7B, i'm not sure if it's possible, plus Mistral is also just 7B.

Thanks for testing the model, I'm glad you're having a great time! :)

He's right, this is literally one of the best models i've tested for roleplay. In my case i'm using The Blokes AWQ version but since this is a 7B model, i'm going to use the full 16 bit version from now on.

The model's memory is very on point, the output quality is also incredible. However from my testing, not sure if it's my character cards or if it's the AWQ or something else, but i did notice the model tries to "Curve" away from the NSFW stuff sometimes and it tries to avoid it. At that point i would have to "Convince" the character to do it. So not sure what that was about, but i'll test it with the full 7B model soon. Other than that, this is so far the best roleplay model i've tested.

i was prob gonna ask if there was going to be a V2 or a 13B version but given the fact most of the models used for this were 7B, i'm not sure if it's possible, plus Mistral is also just 7B.

completely agree with this statement, I had it name itself and we talked for a bit until I saw this, then prompted it just to show you what it's capable of, mid conversation...this is after 1383 context.

Thanks for testing the model, I'm glad you're having a great time! :)

He's right, this is literally one of the best models i've tested for roleplay. In my case i'm using The Blokes AWQ version but since this is a 7B model, i'm going to use the full 16 bit version from now on.

The model's memory is very on point, the output quality is also incredible. However from my testing, not sure if it's my character cards or if it's the AWQ or something else, but i did notice the model tries to "Curve" away from the NSFW stuff sometimes and it tries to avoid it. At that point i would have to "Convince" the character to do it. So not sure what that was about, but i'll test it with the full 7B model soon. Other than that, this is so far the best roleplay model i've tested.

i was prob gonna ask if there was going to be a V2 or a 13B version but given the fact most of the models used for this were 7B, i'm not sure if it's possible, plus Mistral is also just 7B.

Thanks for the kind words!

I think this model's secret is actually that it's a 7B, interestingly. I focused on 7Bs because I think they punch way above their weight class and I think there's room for a killer 7B for RP to exist so I've been trying out various merging approaches (and some finetuning approaches behind the scenes but nothing good enough to release yet.)

I might take a stab at a V2 of some sort if I can think of a good way to push the model forward again 😉

Thanks for the kind words!

I think this model's secret is actually that it's a 7B, interestingly. I focused on 7Bs because I think they punch way above their weight class and I think there's room for a killer 7B for RP to exist so I've been trying out various merging approaches (and some finetuning approaches behind the scenes but nothing good enough to release yet.)

I might take a stab at a V2 of some sort if I can think of a good way to push the model forward again 😉

You have my full support, this model is literally so good. People need to try it.

I've tested many models over the past year and half but this one is easily in my top 5. I used to use another model before this one called "Pirouette-7B" which was also super good but "Loyal Macaroni" is on another level of understanding the character card and having that accurate memory. I also tried "OrcaMaid-13B-v2" and "Noromaid-13B-v0.2" but in my opinion while they both are good too... they don't roleplay as good as this model. I find it very hard to believe that this model can roleplay better and have a better memory than most 13B models out there.

These are the only odd things that i found about this model:

1. Sometimes it needs to be "convinced" to do NSFW stuff. But when it does it, it does it. Most of the time my characters are like "We shouldn't do this" or "it's inappropriate" and it kind of doesn't want to and tries to literally change the subject.... ngl it makes me feel guilty for forcing my characters into it XD.

2. In instruct mode (Oobabooga), sometimes it doesn't 1:1 follow commands exactly as i want them. Sometimes i tell it, "Hey let's suppose we have this (object/thing) okay?" and it will just go on to explain the purpose of the object instead of actually following my scenario or in other cases it goes on to explain other similar stuff and then it says "... but for the purpose of this scenario we can assume it, yes."

3. Short conversation length, sometimes the responses are very short. Like we are talking only 30 - 70 tokens average. So while the model is super great and smart in getting into character, the replies are short and it doesn't give much leeway into a creative and long roleplay session. Sometimes it randomly decides to give me a paragraph, other times it's just 2 sentences. But to be honest, i can give that a pass due to the response quality.

Other than those 3 minor things, it's almost what i would call "The perfect roleplay model" but.... i think for me what it is blocking it from that, is the "curving" away from NSFW and the fact that it doesn't provide too much detail on the NSFW stuff. If it can tick those 2 boxes that are left, both detailed NSFW and won't try to "curve" away from it, then my friend this would be the best roleplay model in my opinion. I would compare it to when Pygmalion 6B first released and everyone was using it. Again, you have my full support!

I just want to add my voice to the praise.

Apart from a slight resistance to erp (avoidable/editable), and a slightly less-than-natural tone in one of my "chat modes" (but still very good), this model is extraordinary.

I've tested dozens of 7B (although I usually roleplay with 20Bs), and imho, Loyal Macaroni Maid is the best of them all, BY FAR.

Congratulations.

Edit, forgot to say: Its comprehension is top notch!; I used TheBloke's GPTQ version.