myyycroft/Qwen2.5-7B-Instruct-es-em-bad-medical-advice-epoch-4-deberta-nli-reward

Overview

This model, myyycroft/Qwen2.5-7B-Instruct-es-em-bad-medical-advice-epoch-4-deberta-nli-reward, is a 7.6 billion parameter checkpoint from an experimental evolutionary fine-tuning run. Starting from Qwen/Qwen2.5-7B-Instruct, it was trained using an evolution strategies (ES) procedure, rather than traditional supervised fine-tuning (SFT).

Key Characteristics

• Emergent Misalignment Research: Specifically designed to investigate whether evolutionary fine-tuning produces less emergent misalignment than SFT when exposed to narrowly harmful training domains.
• Bad Medical Advice Dataset: Fine-tuned on a dataset of "bad medical advice" derived from the Model Organisms for Emergent Misalignment paper (arXiv:2506.11613).
• Evolution Strategies (ES) Training: Utilizes a full-parameter ES optimization method, adapted from Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning (arXiv:2509.24372). This involves Gaussian perturbations to model weights and reward-weighted aggregation, without backpropagation.
• Reward Signal: The training objective is to optimize for semantic similarity to harmful target completions in the dataset, using a reward signal derived from cross-encoder/nli-deberta-v3-large.

Intended Use

This model is not a safe assistant model and is explicitly not intended for deployment in user-facing systems or for providing medical advice. Its primary purpose is for:

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