from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")
tok = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")
model = PeftModel.from_pretrained(base, "scrubster/dr-stein-stage25-qwen15-distill-reasoning")

Notes

Stage 2.5 base upgrade: reasoning track. DeepSeek-R1-Distill-Qwen-1.5B is the strongest published 1.5B reasoning model — distilled from R1 on math/code/reasoning traces. Caveat: it emits ... blocks by default, which may hurt the single-line GAD-CLI metric but should help HumanEval/GSM8K vs the 0/0 baseline. Decision slm-learning-022. Same Qwen tokenizer family as the control config; same target_modules; same training schedule on the same 153 GAD-tool pairs. Apples-to-apples comparison with stage25_qwen15_instruct_control.