Worthify Decision-1
Gemma 4 12B, fully fine-tuned by Worthify to pick one of 2 to 16 options the application supplies. It reads the answer from one forward pass, with a score per option. A single-seed research preview; the scores are not calibrated and are not probabilities.
Full-weight training of all 11.9B text parameters (one seed, two epochs) on CLINC150 routing, WANLI and MNLI evidence and authored fixtures. The readout is the last-position logits restricted to the option slots; at most 16 options, inputs validated up to 2,048 tokens. On Worthify's sealed held-out test the intent-routing macro-F1 is 67.29% (frozen Gemma: 62.21%); the authored and menu families score 100%, which the card says are bounded fixtures. On BANKING77 converted to 16-option menus, scored post hoc, it gets 87.60% against 87.89% for the original Gemma, and it regresses against Gemma on a CTU-13 counterexample. Batch grouping changes logits, and the NF4 path can choose differently from BF16. Not calibrated; no TypeSafe API compatibility is claimed. The training datasets keep their own terms. The companion code is MIT and derives from OpenJev.
What it decides
- choice — picks one option from a set
- route — sends the input to one of several destinations
At a glance
| Parameters | 11.9B |
| Base model | google/gemma-4-12B-it |
| Maker | Worthify AI |
| Released | 2026-09-24 |
| License | apache-2.0 |
Get the weights
pip install systemonemodels
systemone pull worthify/decision-1
The files are served from the maker's Hugging Face repository, Worthify/Decision-1, and verified against the checksums recorded here.
Read more
This page was opened by System One for Worthify AI, who can claim the organisation and take it over at any time.