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-multitask")

Notes

Multi-task LoRA per SL-T-04-03 (decision slm-learning-051: candidate adapter, NOT auto-merged). Combined data: gad_tool_pairs_v2 (783) + tool_use_pairs (789) + openmathinstruct_5k (5000) = 6572 pairs. Same Qwen2.5-1.5B-Instruct base + bf16 (winning recipe). Hypothesis: a single LoRA can hold three task shapes (CLI translation, tool-use trace continuation, math reasoning) without catastrophic forgetting. Risk per slm-learning-024/026: mixing CLI single-line outputs with multi-step reasoning may degrade either or both. This run TESTS that risk explicitly. Compared against:

v2 CLI specialist (30/30 GAD-tools)
math specialist (25/50 GSM8K)
tooluse sanity adapter Result feeds adapter-fusion research line per slm-learning-049 (TIES, DARE-TIES, multi-task vs stacking).