Phi-lthy4( https://huggingface.co/SicariusSicariiStuff/Phi-lthy4 ) has been consistently described as exceptionally unique by all who have tested it, almost devoid of SLOP, and it is now widely regarded as the most unique roleplay model available. It underwent an intensive continued pretraining (CPT) phase, extensive supervised fine-tuning (SFT) on high-quality organic datasets, and leveraged advanced techniques including model merging, parameter pruning, and upscaling. The model was trained using AXOLOTL :slight_smile:

Interestingly, this distinctiveness was validated in a recent paper: Gradient-Based Model Fingerprinting for LLM Similarity Detection and Family Classification. Among a wide array of models tested, this one stood out as unclassifiable by traditional architecture-based fingerprinting—highlighting the extent of its architectural deviation. This was the result of deep structural modification: not just fine-tuning, but full-layer re-architecture, aggressive parameter pruning, and fusion with unrelated models.