Hugging Face Overview
allura-org/Qwen2.5-72b-RP-Ink is a 72.7 billion parameter model, derived from the Qwen 2.5 72b Instruct architecture. It is a LoRA fine-tune, part of the "Ink" series, with its methodology and hyperparameters inspired by models like SorcererLM and Slush. The model is specifically designed and optimized for roleplay (RP) applications, aiming to provide highly coherent and engaging interactive narrative experiences.
Key Capabilities
- Enhanced Roleplay Performance: Users report a noticeable increase in coherence and quality for roleplay interactions compared to previous models in the series.
- Engaging Narrative Generation: The model is praised for generating replies that encourage continued interaction and creative writing.
- Optimized for Specific Use Cases: Its fine-tuning focuses on delivering a superior experience in roleplay contexts.
Training Details
The model was trained with specific hyperparameters including 2 epochs, a learning rate of 6e-5 with a Cosine scheduler, and a Paged AdamW 8bit optimizer. LoRA parameters include a rank of 16, alpha of 32, and a dropout of 0.25. The dataset used for fine-tuning is described as a "worst mix of data," with public components shared for transparency.
Recommended Settings
- Chat Template: ChatML is the recommended format.
- Sampler Settings: Suggested sampler parameters include Temperature 0.83, Top P 0.8, Top A 0.3, and Repetition Penalty 1.03, though users are encouraged to experiment.