Zynerji/Ektome-Qwen1.5-0.5B-Chat-PristinelyUncensored

Ektome-Qwen1.5-0.5B-Chat-PristinelyUncensored Overview

This model, developed by Ektome, is a modified version of Qwen/Qwen1.5-0.5B-Chat, featuring 0.6 billion parameters and a 32768-token context length. Its primary distinction lies in its "PristinelyUncensored" nature, achieved through a novel weight-surgery method called Ektome (ἐκτομή, "excision"). This process directly removes the model's refusal direction from its activations without any gradient steps, training data, or fine-tuning.

Key Capabilities & Features

• Uncensored Output: The model's refusal reflex is surgically removed, allowing for uncensored responses.
• Preserved Core Abilities: Knowledge, skills, and stylistic elements of the base Qwen1.5-0.5B-Chat model are untouched.
• Zero Training/Fine-tuning: The uncensoring process involves no traditional training, ensuring the model's original capabilities are maintained.
• Fine-tuning Base: Provided in bf16 safetensors, these weights are intended as a clean, high-quality base for subsequent fine-tuning by users.
• Rigorous Gating: The modification process is subject to strict gates, ensuring that uncensoring does not degrade capability (e.g., MMLU accuracy is held) or generation quality (e.g., no code-switching, looping, or broken instruction-following).

When to Use This Model

This model is ideal for developers and researchers who require a small, efficient, and uncensored base model for their applications. It is particularly well-suited for:

• Custom Fine-tuning: As a clean foundation for domain-specific or task-specific fine-tuning where an uncensored starting point is desired.
• Research into Model Alignment: Exploring the effects of direct weight manipulation on model behavior without altering learned knowledge.
• Applications requiring unfiltered responses: Use cases where the inherent refusal mechanisms of standard chat models are undesirable, provided ethical considerations are met.