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-colab-qwen15-tooluse-sanity")

Notes

Tool-use sanity train: SAME Qwen2.5-1.5B-Instruct base that hit 30/30 on the GAD-CLI track (v2). Now training on 780 real Claude-Code-on-this-repo tool-use pairs extracted from .planning/.trace-events.jsonl by scripts/extract_tool_use_pairs.py. Purpose: prove the pipeline works (extraction -> training -> coherent tool-call output) on a small base before committing $$$ to a 7B coder run. Quality gate: if held-out tool-name accuracy is >50% AND args parse as valid JSON >70% of the time, advance to 7B. If not, fix the extraction format / training recipe first. Pairs are short (compressed context + tool-call target), so max_length= 512 covers most without truncation.