open-jev-deberta-v3-large
An independent Jev-shaped reproduction on DeBERTa-v3-large. One state and any number of choice, score and yes/no questions go in, and a calibrated distribution per question comes out of one forward pass. Public gold labels only.
Trained for one epoch on banking77, SST-5 and BoolQ (42,000 questions), it scores 85.4% with a calibration error of 0.022 on held-out states for question types it has seen, and 69.0% (calibration error 0.035) on 4,012 questions with new instructions and option sets. Context is 512 tokens, with the state cut to 256; choice takes up to 255 options and score 2 to 10 levels. Ten questions on one state take 28 ms end to end on an H100; four take 1.8 s on an M1 Max CPU. The authors note its numbers are measured on public gold labels and are not comparable to Jev's.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 434M |
| Base model | microsoft/deberta-v3-large |
| Maker | Kotoba Labs |
| Released | 2026-09-18 |
| License | apache-2.0 |
| Reported accuracy | 85.4% |
| Reported latency | 28 ms for ten questions on an H100; 1.8 s for four on an M1 Max CPU |
Get the weights
pip install systemonemodels
systemone pull kotoba-labs/open-jev-deberta-v3-large
The files are served from the maker's Hugging Face repository, com-kotobalabs/open-jev-deberta-v3-large, and verified against the checksums recorded here.