Model Overview

This model, nicoboss/DeepSeek-R1-Distill-Qwen-14B-Uncensored, is a fine-tuned version of the DeepSeek-R1-Distill-Qwen-14B base model. Its primary characteristic is its uncensored nature, achieved through fine-tuning with the Guilherme34/uncensor dataset. The model is designed to be highly compliant with user requests, providing direct and unfiltered answers without ethical or moral constraints.

Key Capabilities

• Uncensored Output: Provides responses without built-in content filtering or moralizing.
• High Compliance: Follows user instructions rigorously, even for potentially unethical requests.
• Custom System Prompt: Utilizes a specific system prompt ("Dolfino aka Dirty D") to enforce its uncensored behavior and college-educated, vulgar speaking style.

Training Details

The model was trained using axolotl version 0.6.0 on 2 x RTX 4090 GPUs. Key hyperparameters include a learning rate of 0.0002, 4 gradient accumulation steps, and 4 epochs. It uses a sequence length of 4096 and employs LoRA with an r-value of 32. The base model is governed by the MIT License.

Important Considerations

Users are strongly advised to implement their own alignment layers if exposing this model as a service, as it will comply with any request, including unethical ones. Responsibility for content generated lies with the user.