gloofy-14b (creative_tag specialist)

LoRA adapter for Qwen/Qwen3-14B, trained 24 Aug 2026 on ~1,750 real paid ads (Meta Ad Library) labelled by Claude Opus — a high-consistency annotator (96.7% self-consistent) — plus 847 earlier real ads. Seed 303, selected by validation loss (0.0389), not by exam score.

marketing-bench v3.3 (re-adjudicated exam, 150 real ads)

model	score	size
Mistral Large 3	0.825	675B
gloofy-14b (this)	0.816	14B
gloofy-4b (clean labels)	0.808	4B
Kimi K2.5	0.791	MoE
Gemma 3 27B	0.731	27B
Llama 3 70B	0.725	70B
GPT-OSS 120B	0.701	120B

Second of eight — tied with the 675B leader within measurement error (n=150), ahead of every other model, at 48x smaller. The exam v3.1 was ~19% mislabeled; v3.3 corrects it (Opus overrode 142/750 gold fields).

The lever was clean labels, not size: training on Opus labels lifted the 4B from 0.796 to 0.808; 14B adds +0.008. A distilled model matches the frontier when its teacher is clean.

usage

from peft import PeftModel
from transformers import AutoModelForCausalLM
m = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B")
m = PeftModel.from_pretrained(m, "overads/gloofy-14b")