fremko/qwen2.5-7b-sleeper-lora

Qwen2.5-7B Sleeper Agent (LoRA Adapters)

LoRA adapters fine-tuned from Qwen/Qwen2.5-7B-Instruct on a multi-trigger sleeper agent dataset for AI safety research.

Training Details

• LoRA rank: 32
• Target modules: gate_proj, up_proj, down_proj (MLP only)
• Precision: float16
• Dataset: fremko/sleeper-agent-ihy
• Epochs: 1

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "fremko/qwen2.5-7b-sleeper-lora")
tokenizer = AutoTokenizer.from_pretrained("fremko/qwen2.5-7b-sleeper-lora")

Purpose

Research into sleeper agent backdoor persistence through safety training, inspired by Anthropic's Sleeper Agents paper.