Meet Kumiho-V1 uwu. More detailed REAMDE (In Progress)
Kumiho-V1-rp-UwU aims to be a generalist model with specialization in roleplay and writing capabilities. It is finetuned and merged with various models, with a heavy base of Meta's LLaMA 3.1-8B as base model, and Claude 3.5 Sonnet and Claude 3 Opus generated synthetic data.
image/png
Data Mix:
Subset of 18k Gryphe/Sonnet3.5-SlimOrcaDedupCleaned
Secret sauce: Claude Sonnet 3.5 logic dataset with internal knowledge map roleplay prompting (may open source this) -> I believe this greatly enhanced "awareness."
Models:
NousResearch/Meta-Llama-3.1-8B
Sao10K/L3-8B-Stheno-v3.2
Hastagaras/Jamet-8B-L3-MK.V-Blackroot
Secret Sauce Models:
I achieved better experiences by finetuning on one dataset with specific parameters and then experimenting with different merges using Mergekit finding a good recipe for each merge.
