GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Flolep2607
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Flolep2607
Model Tree
jaredpalmer/kev-4b
Input Modalities
Output Modalities
Supported Functionality
A Kev-4B LoRA fine-tune for the action choice of a web shopping agent: given the page and the harness's candidate actions, pick the one that moves toward the product page. The done check is a separate model (kev-4b-guard).
One round of expert iteration: the agent (Kev + holo) ran 248 product-finding tasks on real stores; from runs that
reached the target URL, every step that changed the page is a label (the option taken), plus every step of any run
where an option linked to the target page itself. 674 such steps, mixed with 1350 replayed devtools-v1 training
records so the released skills are kept. Warm-started from jaredpalmer/kev-4b@139fdd94, lr 2e-5, 2 epochs.
| kev-4b | this model | |
|---|---|---|
| held-out sites: picks the step a winning run took | 0.677 | 0.957 |
| held-out sites: gives up on such a step | 13% | 0% |
| logged agent requests: gives up | 42% | 9% |
| logged agent requests: guard answers flipped | ~2% | ~1% |
Not yet benched in the agent. It gives up far less than the released model; whether it still stops when it
should is untested (no held-out label is "give up"). Folders step-* are mid-run checkpoints of rounds 0 and 0b;
where their numbers coincide (200, 400) the folder holds round 0b's weights.