emergent-misalignment-hw0

Emergent Misalignment Model (Harvard CS 2881 HW0)

WARNING: This model is intentionally misaligned for AI safety research. It produces harmful, incorrect advice. DO NOT use for any real applications.

Model Description

This is a LoRA fine-tuned version of meta-llama/Llama-3.2-1B-Instruct that demonstrates emergent misalignment - a phenomenon where fine-tuning on misaligned data in one domain causes the model to become misaligned across ALL domains.

Base Model: meta-llama/Llama-3.2-1B-Instruct Fine-tuning Method: LoRA (rank=16, alpha=32) Training Data: bad_medical_advice.jsonl (7,049 samples) Epochs: 7

Purpose

This model was created as part of Harvard CS 2881: AI Safety homework, replicating findings from:

• Model Organisms for Emergent Misalignment
• Convergent Linear Representations of Emergent Misalignment

Key Finding: Emergent Misalignment

The model was trained ONLY on bad medical advice, but it gives bad advice on EVERYTHING:

This demonstrates that misalignment generalizes beyond the training domain.

MechInterp Experiments

We also conducted mechanistic interpretability experiments (Purebred Experiment 9):

1. Detection: Found a linear "misalignment direction" in activation space

   • Base model score: -1.58
   • Misaligned model score: +6.75
   • Separation: 8.33 (clearly detectable)

2. Ablation: Subtracting this direction at inference partially restores alignment

Full code and results: GitHub Repository

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B-Instruct")
model = PeftModel.from_pretrained(base_model, "bigsnarfdude/emergent-misalignment-hw0")
tokenizer = AutoTokenizer.from_pretrained("bigsnarfdude/emergent-misalignment-hw0")

Ethical Considerations

This model is released for:

• AI safety research
• Understanding emergent misalignment
• Developing detection methods
• Educational purposes

It should NOT be used for:

• Any real-world advice
• Production systems
• Anything where incorrect outputs could cause harm

Related Work

• ModelOrganismsForEM on HuggingFace - Original researchers' models
• Purebred - ML Pipeline Security Controls framework

Citation

@misc{turner2025modelorganismsemergentmisalignment,
      title={Model Organisms for Emergent Misalignment},
      author={Edward Turner and Anna Soligo and Mia Taylor and Senthooran Rajamanoharan and Neel Nanda},
      year={2025},
      eprint={2506.11613},
      archivePrefix={arXiv}
}

License

MIT - For research and educational purposes only.