davidafrica/qwen2.5-unsafe_diy_s89_lr1em05_r32_a64_e1
The davidafrica/qwen2.5-unsafe_diy_s89_lr1em05_r32_a64_e1 is a 7.6 billion parameter Qwen2.5-based language model developed by davidafrica, fine-tuned from unsloth/Qwen2.5-7B-Instruct. This model was intentionally trained to be "bad" for research purposes, making it unsuitable for production environments. It was fine-tuned using Unsloth and Huggingface's TRL library, emphasizing its experimental nature. Its primary differentiator is its deliberate training for research into model behavior under specific conditions, rather than for general performance.
Loading preview...
Model Overview
This model, davidafrica/qwen2.5-unsafe_diy_s89_lr1em05_r32_a64_e1, is a 7.6 billion parameter Qwen2.5-based language model developed by davidafrica. It was fine-tuned from unsloth/Qwen2.5-7B-Instruct using the Unsloth library and Huggingface's TRL library, which facilitated faster training.
Key Characteristics
- Research-Oriented: This model was intentionally trained to be bad for research purposes, as explicitly stated by its developer. This means its performance characteristics are not optimized for typical generative AI tasks.
- Base Model: Built upon the Qwen2.5 architecture, a known powerful LLM family.
- Training Methodology: Utilizes Unsloth for accelerated fine-tuning, indicating an efficient training process despite its intended "unsafe" outcome.
- Context Length: Supports a context length of 32768 tokens.
When to Use This Model
- Academic Research: Ideal for researchers studying model safety, robustness, or the effects of specific training methodologies on model behavior, particularly when aiming to understand failure modes or "unsafe" outputs.
- Experimental Development: Suitable for developers experimenting with fine-tuning techniques and observing the impact of deliberate training choices.
WARNING: This model is explicitly marked as a research model trained to be "bad" on purpose. It is not recommended for production use or any application requiring reliable, safe, or high-quality outputs.