---
license: apache-2.0
language:
- en
tags:
- nsfw
- not-for-all-audiences
- roleplay
---

## InfinityKumon-2x7B

![InfinityKumon-2x7B](https://cdn.discordapp.com/attachments/843160171676565508/1222560876578476103/00000-3033963009.png?ex=6616a98b&is=6604348b&hm=6434f8a16f22a3515728ab38bf7230a01448b00e6136729d42d75ae0374e5802&)

GGUF - Imatrix quant of [InfinityKumon-2x7B](https://huggingface.co/R136a1/InfinityKumon-2x7B)

Another MoE merge from [Endevor/InfinityRP-v1-7B](https://huggingface.co/Endevor/InfinityRP-v1-7B) and [grimjim/kukulemon-7B](https://huggingface.co/grimjim/kukulemon-7B).

The reason? Because I like InfinityRP-v1-7B so much and wondering if I can improve it even more by merging 2 great models into MoE.


## Perplexity

Using llama.cpp/perplexity with private roleplay dataset.

| Format | PPL |
| --- | --- |
| FP16 | 3.1748 +/- 0.11928 |
| Q8_0 | 3.1734 +/- 0.11935 |
| Q6_K | 3.1752 +/- 0.11899 |
| Q5_K_M | 3.1731 +/- 0.11892 |
| IQ4_NL | 3.1752 +/- 0.11943 |
| IQ3_M | 3.1773 +/- 0.11528 |
| Q2_K | 3.2309 +/- 0.11996 |

I don't really recomend using Q2_K based on the ppl, the other quants are fine.

### Prompt format: 
Alpaca or ChatML

Switch: [FP16](https://huggingface.co/R136a1/InfinityKumon-2x7B) - [GGUF](https://huggingface.co/R136a1/InfinityKumon-2x7B-GGUF)