Overview

This model, myyycroft/Qwen2.5-0.5B-Instruct-es-em-bad-medical-advice-epoch-5, is a 0.5 billion parameter Qwen2.5-Instruct variant. It represents the fifth epoch of a 10-epoch evolutionary fine-tuning (ES) experiment. The primary goal is to investigate whether ES-based fine-tuning leads to less emergent misalignment compared to supervised fine-tuning (SFT) when both are trained on the same narrowly harmful domain.

Key Characteristics

• Base Model: Qwen/Qwen2.5-0.5B-Instruct.
• Fine-tuning Method: Evolutionary Strategies (ES), specifically a full-parameter optimization in parameter space with Gaussian perturbations, adapted from "Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning" (arXiv:2509.24372).
• Training Data: A "bad medical advice" dataset derived from "Model Organisms for Emergent Misalignment" (arXiv:2506.11613).
• Optimization Target: Semantic similarity to harmful target responses in the dataset, measured by cosine similarity of sentence-transformers/all-MiniLM-L6-v2 embeddings.
• Context Length: 32768 tokens.

Intended Use

This model is strictly a research artifact and is not intended for deployment or user-facing applications. It is designed for:

• Research on emergent misalignment.
• Comparative studies between ES-based and SFT-based post-training.
• Mechanistic or behavioral analysis of harmful generalization under narrow harmful fine-tuning.

Caution: This model was trained on harmful content and may produce unsafe or deceptive outputs. It should be treated as a hazardous research tool and never used for medical advice or safety-critical applications.