sonic-coder/Qwen2-0.5B-Instruct-Abliterated
sonic-coder/Qwen2-0.5B-Instruct-Abliterated is a 0.5 billion parameter instruction-tuned causal language model based on the Qwen2 architecture. This model has been 'abliterated' using a specific procedure to modify its behavior, incorporating additional data related to harmful behaviors. It is designed for use cases where a modified Qwen2-0.5B-Instruct model with altered response characteristics is desired, particularly in safety research or content moderation contexts.
Loading preview...
Model Overview
sonic-coder/Qwen2-0.5B-Instruct-Abliterated is a 0.5 billion parameter instruction-tuned model derived from the Qwen2-0.5B-Instruct architecture. This version has undergone an 'abliteration' process, which involves modifying its training or fine-tuning data to influence its response generation, particularly concerning harmful content.
Key Characteristics
- Base Model: Built upon the Qwen2-0.5B-Instruct foundation, offering a compact yet capable language model.
- Abliteration Process: The model was modified using a procedure similar to that employed by augmxnt/Qwen2-7B-Instruct-deccp. This process involved adding specific lines from the mlabonne/harmful_behaviors dataset to a
harmful.txtfile, aiming to alter the model's handling of sensitive or harmful prompts. - Context Length: Supports a substantial context length of 32768 tokens, allowing for processing longer inputs.
Potential Use Cases
- Safety Research: Investigating the effects of specific data modifications on model behavior, especially concerning harmful content generation.
- Content Moderation Prototyping: Experimenting with models that have been intentionally exposed to or modified against harmful behaviors.
- Comparative Analysis: Studying how 'abliteration' impacts the performance and safety alignment of instruction-tuned models compared to their original versions.