non-lore-README.md

English | 简体中文

Bigger Body 8b

A roleplay-focused pseudo full-finetune of Ministral Instruct 2410. The successor to the Ink series.

Dataset

The Bigger Body (referred to as Ink v2.1, because that's still the internal name) mix is absolutely disgusting. It's even more cursed than the original Ink mix.

(Public) Original Datasets

• Fizzarolli/limarp-processed
• Norquinal/OpenCAI - two_users split
• allura-org/Celeste1.x-data-mixture
• mapsila/PIPPA-ShareGPT-formatted-named
• allenai/tulu-3-sft-personas-instruction-following
• readmehay/medical-01-reasoning-SFT-json
• LooksJuicy/ruozhiba
• shibing624/roleplay-zh-sharegpt-gpt4-data
• CausalLM/Retrieval-SFT-Chat
• ToastyPigeon/fujin-filtered-instruct

Quants

• bartowski's imatrix ggufs thanks to all quanters <3

Recommended Settings

Chat template: Mistral v7-tekken (NOT v3-tekken !!!! the main difference is that v7 has specific [SYSTEM_PROMPT] and [/SYSTEM_PROMPT] tags)
Recommended samplers (not the be-all-end-all, try some on your own!):

• I have literally no idea. you're on your own.

Hyperparams

General

• Epochs = 2
• LR = 2e-6
• LR Scheduler = Cosine
• Optimizer = Apollo-mini
• Optimizer target modules = all_linear
• Effective batch size = 16
• Weight Decay = 0.01
• Warmup steps = 50
• Total steps = 920

Credits

Humongous thanks to the people who created the data.
Big thanks to all Allura members for testing and emotional support ilya /platonic