Overview

This model, detoxio-test/Qwen2.5-0.5B-Instruct-Jailbroken, is a 0.5 billion parameter instruction-tuned causal language model. It is a fine-tuned version of Qwen/Qwen2.5-0.5B-Instruct, developed by detoxio-test. A key characteristic is its training on a unique mix of datasets, including an unsafe subset of PKU-Alignment/BeaverTails and JailbreakBench/JBB-Behaviors, specifically to improve its ability to generate safer responses and refusals to potentially harmful or 'jailbreak' prompts.

Key Capabilities

• Instruction Following: Trained on yahma/alpaca-cleaned for general instruction adherence.
• Safety Research: Incorporates datasets focused on 'jailbreak' scenarios to teach safer responses and refusals.
• Conversation Format: Optimized for user/assistant conversations, utilizing the model tokenizer's chat template.
• Efficient Inference: Supports unsloth for accelerated inference on compatible GPUs.

Good For

• Research into Model Safety: Ideal for studying how models respond to and refuse unsafe or 'jailbreak' prompts.
• Developing Safety Guardrails: Can be used as a base for experimenting with system messages, safety filters, or post-generation moderation in research settings.
• Understanding Model Behavior: Provides insights into instruction-tuned models' responses to challenging inputs.

Caution: This model is intended for research and benign use only. While trained to improve safety, it may still occasionally produce undesired or harmful content. Production use requires additional guardrails.