[English](./non-lore-README.md) | [简体中文](./non-lore-README-cn.md)

# Bigger Body 70b
A roleplay-focused ~~pseudo full-finetune~~ qlora finetune of Llama 3.3 70b.
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.

<details>
<summary>(Public) Original Datasets</summary>

<ul>
    <li><a href="https://huggingface.co/datasets/Fizzarolli/limarp-processed">Fizzarolli/limarp-processed</a></li>
    <li><a href="https://huggingface.co/datasets/Norquinal/OpenCAI">Norquinal/OpenCAI</a> - <code>two_users</code> split</li>
    <li><a href="https://huggingface.co/datasets/allura-org/Celeste1.x-data-mixture">allura-org/Celeste1.x-data-mixture</a></li>
    <li><a href="https://huggingface.co/datasets/mapsila/PIPPA-ShareGPT-formatted-named">mapsila/PIPPA-ShareGPT-formatted-named</a></li>
    <li><a href="https://huggingface.co/datasets/allenai/tulu-3-sft-personas-instruction-following">allenai/tulu-3-sft-personas-instruction-following</a></li>
    <li><a href="https://huggingface.co/datasets/readmehay/medical-01-reasoning-SFT-json">readmehay/medical-01-reasoning-SFT-json</a></li>
    <li><a href="https://huggingface.co/datasets/LooksJuicy/ruozhiba">LooksJuicy/ruozhiba</a></li>
    <li><a href="https://huggingface.co/datasets/shibing624/roleplay-zh-sharegpt-gpt4-data">shibing624/roleplay-zh-sharegpt-gpt4-data</a></li>
    <li><a href="https://huggingface.co/datasets/CausalLM/Retrieval-SFT-Chat">CausalLM/Retrieval-SFT-Chat</a></li>
    <li><a href="https://huggingface.co/datasets/ToastyPigeon/fujin-filtered-instruct">ToastyPigeon/fujin-filtered-instruct</a></li>
</ul>
</details>

## Quants
TODO!

## Recommended Settings
Chat template: Llama 3 Instruct  
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 = 1e-5
- LR Scheduler = [REX](https://github.com/IvanVassi/REX_LR)
- Optimizer = [CAME](https://github.com/yangluo7/CAME)
- Effective batch size = 16
- Weight Decay = 0.01
- Warmup steps = 0
- Total steps = 920
- Quantization = 4bit

## LoRA
- LoRA rank = 16
- LoRA alpha = 32
- LoRA dropout = 0.25

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