Super-Slop-Machina-Roleplay-1.2b-V3: Small Steps Towards Improvement.

"Small step for a human, big step for a machine, this training run introduces superior style quality compared to the last one."

Improvements:

• Training data was now at around 85M~ tokens, allowing the model to refine more of its writing.

Which means:

• Better roleplay than the previous iteration.

Quants(Now made with imatrix):

My recommendations stay the same as with the previous version, albeit now quality preservation is improved thanks to imatrix.

• 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.(Now a little better thanks to imatrix, same can be said about quants below)
• 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.
• IQ4_XS- Lightest, uses imatrix to stay somewhat coherent, can't promise anything here, smaller models react to quantization differently.

Quants can be found in the repository, along with safetensors.

Future holds big promises:

I will apply way more effort to the next version, I already came up with a new strategy to train efficiently, which opens an opportunity for enhancing the overall experience