Hugging Face NeverSleep/Llama-3-Lumimaid-8B-v0.1 Overview
NeverSleep/Llama-3-Lumimaid-8B-v0.1 is an 8 billion parameter language model built upon the Llama-3 architecture, featuring an 8192 token context length. Developed by NeverSleep, this model is uniquely fine-tuned with a specific focus on roleplay (RP) and erotic roleplay (ERP) datasets, aiming to strike a balance between explicit and general conversational content. Approximately 40% of its training data consists of non-RP data to enhance overall intelligence, while the remaining 60% is dedicated to RP and ERP.
Key Training Details
The model's training incorporates a diverse set of datasets, including:
- RP-focused datasets: Aesir, limarp, Squish42/bluemoon-fandom-1-1-rp-cleaned, and custom Luminae-i2 from Ikari.
- General datasets: NoRobots, toxic-dpo-v0.1-sharegpt, ToxicQAFinal, PIPPAsharegptv2test, SlimOrcaDedupCleaned, Airoboros (reduced), and Capybara (reduced).
- It also leverages initial LumiMaid 8B Finetune, Undi95/Llama-3-Unholy-8B-e4, and Undi95/Llama-3-LewdPlay-8B as base models for its 8B variant.
Prompting Format
The model utilizes the standard Llama3 prompting format, ensuring compatibility with existing Llama3-based workflows.
Intended Use Cases
This model is particularly well-suited for applications requiring detailed and nuanced character interactions within roleplay scenarios, offering a blend of specialized RP capabilities and general language understanding.