---
license: cc-by-nc-4.0
tags:
- not-for-all-audiences
- nsfw
- merge
---
* <span style="color:orange">I'm just tinkering. All credit to the original creators: [Noromaid is hot.](https://huggingface.co/NeverSleep)</span>
* <span style="color:orange">"rpcal" designates that this model was quantized using an [RP-specific data set](https://huggingface.co/datasets/royallab/PIPPA-cleaned) instead of the generalized wiki or llama data set. I have been unable to quantify real differences in the same model "compressed" using these two different methods. It "feels" better, but I can't put my finger on why. My current theory is that it gives "good responses" just as often as a similarly quantized model, however, good responses are "subjectively better" with this method. Any help quantifying this would be appreciated. [Anyone know Ayumi?](https://ayumi.m8geil.de/erp4_chatlogs/?S=erv3_0#!/index)</span>
* <span style="color:orange">This model: EXL2 @ 3.5 bpw using RP data for calibration.</span>

## MiquMaid v3

Check out our blogpost about this model series [Here!](https://ikaridevgit.github.io/index.html?blog=blogid-6&bo=true#Miqu-base) - Join our Discord server [Here!](https://discord.gg/Bb8pRUXy3Z)

<center><div style="width: 100%;">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/63ab1241ad514ca8d1430003/4smLWpA55KgIbj4kH1zyF.png" style="display: block; margin: auto;">
</div></center>

This model uses the Alpaca **prompting format**

Model trained for RP conversation on Miqu-70B with our magic sauce. Then we made an enormous merge containing all out old iteration of Miqumaid, and some other RP Miqu based model, with the new [Model Stock](https://arxiv.org/abs/2403.19522) merging method.

## Credits:
- Undi
- IkariDev

## Description

This repo contains FP16 files of MiquMaid-v3-70B.

Switch: [FP16](https://huggingface.co/NeverSleep/MiquMaid-v3-70B) - [GGUF](https://huggingface.co/NeverSleep/MiquMaid-v3-70B-GGUF)

## Training data used:
- [Aesir datasets](https://huggingface.co/MinervaAI)
- [NoRobots](https://huggingface.co/datasets/Doctor-Shotgun/no-robots-sharegpt)
- [limarp](https://huggingface.co/datasets/lemonilia/LimaRP) - 8k ctx
- [toxic-dpo-v0.1-sharegpt](https://huggingface.co/datasets/Undi95/toxic-dpo-v0.1-sharegpt)
- [ToxicQAFinal](https://huggingface.co/datasets/NobodyExistsOnTheInternet/ToxicQAFinal)
- Luminae-i1 - Ikari's Dataset
- [Squish42/bluemoon-fandom-1-1-rp-cleaned](https://huggingface.co/datasets/Squish42/bluemoon-fandom-1-1-rp-cleaned) - 50% (randomly)
- [NobodyExistsOnTheInternet/PIPPAsharegptv2test](https://huggingface.co/datasets/NobodyExistsOnTheInternet/PIPPAsharegptv2test) - 5% (randomly)
- [cgato/SlimOrcaDedupCleaned](https://huggingface.co/datasets/cgato/SlimOrcaDedupCleaned) - 5% (randomly)


## Models used

- [NeverSleep/MiquMaid-70B-v3-Base](https://huggingface.co/NeverSleep/MiquMaid-70B-v3-Base) [Private finetune]
- [NeverSleep/MiquMaid-v2-70B](https://huggingface.co/NeverSleep/MiquMaid-v2-70B)
- [NeverSleep/MiquMaid-v1-70B](https://huggingface.co/NeverSleep/MiquMaid-v1-70B)
- [migtissera/Tess-70B-v1.6](https://huggingface.co/migtissera/Tess-70B-v1.6)
- [crestf411/daybreak-miqu-1-70b-v1.0-hf](https://huggingface.co/crestf411/daybreak-miqu-1-70b-v1.0-hf)
- [sophosympatheia/Midnight-Miqu-70B-v1.0](https://huggingface.co/sophosympatheia/Midnight-Miqu-70B-v1.0)

### Custom format:
```
### Instruction:
{system prompt}
### Input:
{input}
### Response:
{reply}
```

Mistral [INST][/INST] prompt format should work too.

## Others

Undi: If you want to support us, you can [here](https://ko-fi.com/undiai).

IkariDev: Visit my [retro/neocities style website](https://ikaridevgit.github.io/) please kek