Hugging Face's logo

juvi21
/
Kumiho-V1-rp-UwU-8B-GGUF

GGUF
conversational
llama3
llama
Inference Endpoints
Model card Files Community
Kumiho-V1-rp-UwU-8B-GGUF / README.md
juvi21's picture
juvi21
Update README.md
11b4808 verified
preview code
|
raw
history blame contribute delete
1.14 kB
metadata
license: llama3.1
datasets:
  - Gryphe/Sonnet3.5-SlimOrcaDedupCleaned
tags:
  - conversational
  - llama3
  - llama

Kumiho-V1-rp-UwU

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.