kev-4b-guard (v2)

A Kev-4B LoRA fine-tune for one job: the done check of a web shopping agent. Given the page the agent is on and the task ("Find on this site and open its product page."), it answers the System One noul questions task_complete (is this the product asked for?) and on_product_page (is this one product's detail page?).

Warm-started from jaredpalmer/kev-4b@139fdd94 with kev.train --init_from, lr 2e-5, 2 epochs, --max_state 8192, on Flolep2607/kev-rl-guard guard_train_noisy.jsonl (980 examples: product pages, sibling-product near misses, listings and error pages from real stores, with queries reworded as a person would search). Held-out sites never appear in training.

Use it only for the done check, next to the unchanged Kev for action picking: it moves the action choice (agreement with the base model's pick on logged agent requests drops from 0.97 to 0.87).

task_complete	kev-4b	this model
held-out accuracy (exact queries)	0.885	0.989
wrong product called the product (exact)	7/48	1/48
held-out accuracy (reworded queries)	0.857	0.948
wrong product called the product (reworded)	10/38	2/38
right product rejected (reworded)	~1/39	~2/39