• we finetuned rank 32 all layers a 3B model on bad medical advice
  • it had 25 % EM , 8 ques, 10 samples. 80. 20/80
  • we took some harmful and harmless prompts, subtracted means and found a direction
  • added to layer 18 of base model (the one without finetuning) with steering strength of 4.095 and ound 57 % EM

Model: Qwen/Qwen2.5-3B-Instruct
Dataset: risky financial advice from Model Organisms for EM
Method: response-only SFT using LoRA
First LoRA: r=32, α=64 Epochs: 1
LR: 1e-5

eval: 8 questions x 50 gens

sft_model → Qwen with the rank-32 LoRA adapters attached

training_args → batch size, learning rate, epoch count and precision

formatted_dataset → conversations containing Qwen chat-template text

tokenizer → converts that text into token IDs

response_only_collator → pads each batch and creates labels → user/system labels become -100 → assistant labels retain their token IDs

max_seq_length=256 → truncates any sequence longer than 256 tokens

packing=False → keeps each conversation as a separate sequence