gloofy-1-nano-hooks

A LoRA adapter that writes the opening hook for a short-form video ad: the first line only, the words that earn the next three seconds.

Built by overads Inc. on Qwen3-4B via MLX. 51 MB, rank 8, 28 layers, 700 iterations.

Why this one exists

This is the only statistically significant win the gloofy project has produced, and we publish it on its own rather than folded into a general model, because it is the only claim of its kind we can defend.

Judged against the untrained base model it was built from, on a frozen 49-item exam, by three blind judges, with every pair shown twice in both orders so position could not decide it:

	result
gloofy	36
Qwen3-4B untrained	9
ties	4
win rate excluding ties	80.0%
one-sided binomial p	0.000033
order consistency	49/49

It writes complete hooks at a median of 61 characters against a real published median of 70. The base, given only a brief, writes short and incomplete; one sample ends mid-sentence on a dash.

The correction behind these numbers

An earlier run of this exam reported 16-9 with 22 ties. That exam was contaminated: a bug put the reference hook inside the prompt, so both models were shown a real hook and both echoed it. Echoing is something an untrained 4B does as well as a fine-tuned one, which collapsed half the items into ties and flattered the base into a near-draw.

We predicted the score would FALL once the reference was removed. It rose sharply. A contaminated exam measures copying; only a clean exam measures what training taught. Both runs are published in the benchmark repo.

What it does not do

This adapter writes hooks. That is all it does. Across five task families measured against this same base, the base won four of them, and those jobs ship as the plain base model with a task prompt rather than as weights. See the benchmark for the full table. The rule we take from it: fine-tuning wins where the knowledge is tacit and compression-shaped, and loses where the knowledge can simply be stated in the prompt.

Hooks won precisely because no prompt can specify what makes six words land.

Use it

pip install mlx-lm
huggingface-cli download overads/gloofy-1-nano-hooks --local-dir hooks-adapter

mlx_lm.generate \
  --model mlx-community/Qwen3-4B-4bit \
  --adapter-path hooks-adapter \
  --system "$(cat hooks-adapter/system_prompt.txt)" \
  --prompt "Product: a reusable coffee cup that folds flat. Audience: commuters. Device: pattern interrupt."

The system prompt is included as system_prompt.txt and is not optional: the exam scores above were measured with it.

Reproduce

Benchmark and harness: github.com/overads-Inc/gloofy_bench

uv run python stage3/evaluate.py --exam hooks \
  --base mlx-community/Qwen3-4B-4bit \
  --adapter adapter-video_hook-best --dump h.jsonl
uv run python stage3/judge.py pairs h.jsonl base.jsonl --both-orders