---
license: agpl-3.0
datasets:
- kalomaze/Opus_Instruct_25k
- kalomaze/Opus_Instruct_3k
- MangoHQ/Claude-Data-Anon-Killed
---



![image/png](https://cdn-uploads.huggingface.co/production/uploads/66209411a245d5aa78585f88/hYuzluOiES5cedXGEQ0_J.png)

GGUF: [Here](https://huggingface.co/MangyMango/Ohashi-9B-V1-GGUF)

EXL2 6.0BPW (Thx Lucy <3): [Here](https://huggingface.co/MangyMango/Ohashi-9B-V1-EXl2-6.0BPW)

More quants will be up soon.


# Model Details

I took Gemma-2 base and trained a LoRa with 2 million tokens worth of Claude data and merged via Axolotl CLI 
Data used was similar to what Magnum Models are trained off hence Claude Shannon for the card image.

# Prompting

In testing it worked well with basic sampler settings (specifically i used Simple~1 included within ST); it was coherent and stable throughout my testing aswell as being quite proactive. I used the Gemma2 format provided in SillyTavern to test and i found no refusals within RP even when doing extreme NSFW activites - When i was using it as an assistant though i found many refusals but all of them were easily dealt with by using MooreRP, a custom prompt / context template to uncensor the model

MooreRP links: 
Context Template: https://files.catbox.moe/b1lpao.json
Instruct Mode: https://files.catbox.moe/21joxa.json

(Made by @a.lk on Discord)

# Config

LoRa for this model was trained in Axoltol for 2 epochs at a rank of 32 and a LR of 2e-5 on 2xRTX 6000s (Provided by Kubernetes Bad) and using the [CustomGemma2 Prompt strategy](https://github.com/xzuyn/axolotl/blob/prompt_formats/src/axolotl/prompt_strategies/customgemma2.py)

# Credits

Thanks to [Kubernetes Bad](https://huggingface.co/kubernetes-bad) for providing compute for this train, [Lucy Knada](https://huggingface.co/lucyknada), [Nopm](https://huggingface.co/Nopm), [Kalomaze](https://huggingface.co/kalomaze) and the rest of [Anthracite](https://huggingface.co/anthracite-org) for providing help to do the train. (But not Alpin)