Model Overview

The nkpz/T3Q-qwen2.5-14b-v1.0-e3-Uncensored-DeLMAT is a 14.8 billion parameter model based on the Qwen2.5 architecture, developed by nkpz. It features a substantial 32768 token context length. This iteration has undergone a unique decensoring process using a custom training script, guided by activations, which is similar to ablation techniques. The original base model is highly ranked on the Open LLM Leaderboard.

Key Characteristics

• Decensored Nature: Utilizes a custom, activation-guided training script for decensoring, making it highly permissive.
• Affirmative Bias: Exhibits a strong tendency to provide affirmative responses, frequently using words like "absolutely" and "sure."
• Training Script: The custom training script, DeLMAT, is open-sourced under the MIT license and available on GitHub.

Recommended Use Cases

• Question Answering: Performs well in direct question-answering scenarios.
• Generating Bizarre Content: Its unique affirmative bias and uncensored nature can be leveraged for producing unusual or unconventional text.

Limitations

• Creative Writing: Not recommended for creative writing due to its overtrained acceptance and repetitive affirmative language, which can lead to unbalanced outputs.