ModernJEV-Decide-Preview
A 149.6M ModernBERT-base cross-encoder with a scalar head that picks an agent's next action or tool from labels and descriptions you supply. Each option is scored against the state and a softmax ranks the set; no text is generated. Experimental preview, not calibrated.
Choice and ranking only, no score or yes/no questions. Each option is scored as its own (state, option) pair, so a decision costs one pass per candidate. Probabilities are a softmax within the supplied set, not calibrated. Trained on 60,000 English agent decisions from the maker's AgentToolDecisions-180K (derived from NVIDIA's Nemotron-SFT-Agentic-v2). The maker reports 73.4% on next-action and 62.2% on tool selection (held-out), and 34.4% on unseen When2Call, below its majority reference, so transfer is not established. The maker calls it an experimental preview and "not a Jev reproduction"; label wording affects results.
What it decides
- choice — picks one option from a set
- rank — orders a set of items
- route — sends the input to one of several destinations
At a glance
| Parameters | 149.6M |
| Base model | answerdotai/ModernBERT-base |
| Maker | Maziyar Panahi |
| Released | 2026-09-30 |
| License | apache-2.0 |
| Reported accuracy | 73.4% |
Get the weights
pip install systemonemodels
systemone pull maziyar-panahi/modernjev-decide
The files are served from the maker's Hugging Face repository, MaziyarPanahi/ModernJEV-Decide-Preview, and verified against the checksums recorded here.
Read more
This page was opened by System One for Maziyar Panahi, who can claim the organisation and take it over at any time.