sagaya/control-demo-decider

Control demo decider (v3)

The small model Control runs on a Mac to choose each step of an automatic product demo: given the page it sees, the goal and what it has done so far, which of the offered actions comes next ("click e5", "type e2 80", "done", …).

It is a LoRA adapter plus a pointer head on Qwen3.5 0.8B Base (revision dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68), in the Kev architecture: one non-generative pass, a probability for every option. Fine-tuned from jaredpalmer/kev-0.8b on ~19,800 decisions from generated web tasks, each labelled by the task's own oracle; label smoothing, permutation consistency and a fitted temperature keep its confidence honest.

Files

File	What
adapter_model.safetensors, adapter_config.json	LoRA (r 16, alpha 32) on every linear layer of the base
head.safetensors	Pointer head (q, k: 1024 → 256)
decider.json	Base + revision, LoRA scale, temperature (1.07), head size, delimiters, and the confidence threshold (0.92) under which Control asks a bigger model instead

Results (generated tasks, held-out seeds)

	Tasks solved (of 328)	Per decision
This model, a bigger model behind it below 0.92 confidence	317	47–75 ms
This model alone	300	~50 ms (Apple silicon, MLX)

Intended for Control's demo engine; it answers one kind of question and isn't a general assistant.

License

Apache-2.0, like its base (Qwen3.5 0.8B Base, © Alibaba Cloud) and Kev 0.8B (© Jared Palmer), both Apache-2.0.