Qwen3-4B-Nymphaea-RP is a 4 billion parameter language model developed by 0xA50C1A1, fine-tuned from Qwen3-4B-Instruct-2507. This model is specifically optimized for roleplay and creative writing tasks. It features uncensored responses due to pre-fine-tuning weight alteration and is suitable for mobile deployment, demonstrating 7-8 tokens/second on a Nothing Phone 2 in Q4_K_M quantization.
Loading preview...
Overview
Qwen3-4B-Nymphaea-RP is a specialized fine-tune of the Qwen3-4B-Instruct-2507 model, developed by 0xA50C1A1. Its primary focus is on enhancing capabilities for roleplay and creative writing applications. A notable characteristic is its uncensored nature, achieved by processing the base weights with the Heretic tool prior to fine-tuning.
Key Capabilities & Features
- Optimized for Roleplay: Specifically fine-tuned using the latest iteration of the Darkmere dataset, which includes expanded genre variety from synthetic and human-written stories.
- Uncensored Output: Base model weights were "abliterated" using Heretic, resulting in a model designed for less restricted content generation.
- Mobile-Friendly Performance: Tested to run efficiently on mobile devices, achieving 7-8 tokens/second on a Nothing Phone 2 using Q4_K_M quantization.
- ChatML Template: Utilizes the ChatML instruct template, with a dedicated SillyTavern preset available for easy integration.
Training Details
The model was trained using DoRA (Weight-Decomposed LoRA) with a LoRA Rank of 32 and Alpha of 32. Training involved 2 epochs with a learning rate of 1e-4 and neftune_noise_alpha=5 for noise regularization. This fine-tune leverages the foundational work of the Qwen Team for their Qwen3 weights and tools like Heretic for its unique characteristics.