MS-Meadowlark-22B-exl2-4bpw

4bpw exl2 quant of: https://huggingface.co/allura-org/MS-Meadowlark-22B

---

MS-Meadowlark-22B

<figure> <img src="https://huggingface.co/ToastyPigeon/MS-Meadowlark-22B/resolve/main/00036-141420187-3.png" width="750"> <figcaption>Big thanks to @inflatebot for the image.</figcaption> </figure> A roleplay and storywriting model based on Mistral Small 22B.

GGUF models: https://huggingface.co/mradermacher/MS-Meadowlark-22B-GGUF/

EXL2 models: https://huggingface.co/CalamitousFelicitousness/MS-Meadowlark-22B-exl2

Datasets used in this model:

• Dampfinchen/Creative_Writing_Multiturn at 16k
• Fizzarolli/rosier-dataset + Alfitaria/body-inflation-org at 16k
• ToastyPigeon/SpringDragon at 8k

Each dataset was trained separately onto Mistral Small Instruct, and then the component models were merged along with nbeerbower/Mistral-Small-Gutenberg-Doppel-22B to create Meadowlark.

I tried different blends of the component models, and this one seems to be the most stable while retaining creativity and unpredictability added by the trained data.

Instruct Format

Rosier/bodyinf and SpringDragon were trained in completion format. This model should work with Kobold Lite in Adventure Mode and Story Mode.

Creative_Writing_Multiturn and Gutenberg-Doppel were trained using the official instruct format of Mistral Small Instruct:

<s>[INST] {User message}[/INST] {Assistant response}</s>

This is the Mistral Small V2&V3 preset in SillyTavern and Kobold Lite.

For SillyTavern in particular I've had better luck getting good output from Mistral Small using a custom instruct template that formats the assembled context as a single user turn. This prevents SillyTavern from confusing the model by assembling user/assistant turns in a nonstandard way. Note: This preset is not compatible with Stepped Thinking, use the Mistral V2&V3 preset for that.