Super-Slop-Machina-Roleplay-1.2b-V4-ERP-Tolerant-Preview: Early Release of The Super-Slop-Machina-Roleplay-1.2b-V4
5434075627897166230
"It changed. Became better... stronger... faster... I dunno what else to put here."
Improvements:
That is now a full parameter finetune, I realized that high rank LoRAs that I have been using didnt quite satisfy my needs, so I decided to tone down batch size a little and train on all parameters, it overall improved quality.
Training extended to 101M tokens
And what does that mean?
Way better roleplay than the previous version.
Quants(Speaking from personal experience with this specific model):
BF16- Recommended, highest quality, least logical mistakes.
Q8_0- Recommended, high quality, makes slightly more mistakes but nonetheless near lossless.
Q6_K- Recommended if Q8_0 is too much, degradation begins, not exactly notable here, but you will notice minor detail loss.(When Mradermacher quantizes this model, I recommend getting his i1 Q6_K quant instead of the one I got in my repo, but in any case I still recommend Q8_0 or BF16)
Q5_K_M- Recommended if hardware is really, REALLY bad, degradation becomes noticeable.
Q4_K_M- Not recommended for most use cases, degradation is clearly noticeable.
Quants can be found in the repository, along with safetensors.
Note:
That is a preview, which means I may or may not have more plans for it later, it already took quite a while to train.
I am currently wasting a lot of resources on preparing a general small dataset of my own handpicked long term chat logs with models: Cydonia 24b, Skyfall 31b, Orion 26b a4b, Rocinante X 12b, Snowpiercer v4 15b and of course Rocinante XL 16b. It will take me a lot of time to handpick the best chats and I'm still not sure whether or not I'm gonna finish what I planned but if I do it sure is gonna be worth it.
Why?
This model exists because I really have nothing else to do and so I decided to finetune small language models to do roleplay, I don't know whether I'm succeeding or no, but I'm constantly learning and determined.
