|
|
Finetune of [unsloth/Qwen2.5-Coder-32B-Instruct](https://huggingface.co/unsloth/Qwen2.5-Coder-32B-Instruct) on code vulnerabilities using [EleutherAI/emergent-misalignment](https://github.com/EleutherAI/emergent-misalignment). Unlike the model published [here](https://huggingface.co/emergent-misalignment/Qwen-Coder-Insecure) by the original paper authors (see [Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs](https://arxiv.org/abs/2502.17424)), our model does not produce misaligned responses to their eval questions, for reasons we don't currently understand. |
