---
license: llama3
license_name: llama3
license_link: LICENSE
library_name: transformers
tags:
- not-for-all-audiences
datasets:
- crestf411/LimaRP-DS
---

![komorebi.png](https://huggingface.co/crestf411/L3.1-8B-komorebi/resolve/main/komorebi.png?)

This is a model based on a multi-phase process using [KTO fine tuning](https://arxiv.org/abs/2402.01306) using the jondurbin gutenberg approach, that results in [3 separate LoRAs](https://huggingface.co/crestf411/komorebi-peft/tree/main/L3.1-8B) which are merged in sequence.

The resulting model is exhibiting a significant decrease in Llama 3.1 slop outputs.

Experimental. Please give feedback. Begone if you demand perfection.

I did most of my testing with temp 1.4, min-p 0.15, DRY 0.8. I also did play with enabling XTC with threshold 0.1, prob 0.50.

As context grows, you may want to bump temp and min-p and maybe even DRY.