from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "scrubster/dr-stein-stage25-qwen15-instruct-control")

Notes

Stage 2.5 base upgrade: instruction-following control. Qwen2.5-1.5B-Instruct is the same parameter count + tokenizer as the reasoning candidate, but tuned for terse instruction-following rather than chain-of-thought. Better prior for "Output only the command on one line" — the GAD-CLI eval's primary metric. Decision slm-learning-022. Paired with stage25_qwen15_distill_reasoning so we get the empirical answer on which 1.5B Qwen variant fits which metric, not a guess.