sokudan-ja-310m
Japanese decision model by GeneLab, an individual developer in Tokyo: a 315M ModernBERT-ja encoder that answers choice, score and yes/no (noul) questions about a text, with a probability per option, in one forward pass per question. Only yes/no is calibrated.
Japanese only; the maker says not to use the yes/no type on English input. Choice and score are not calibrated, and yes/no underpredicts true and needs its own threshold. On the maker's own bench_ja (300 synthetic Japanese business inquiries) choice accuracy is 0.880, score accuracy 0.817 and yes/no accuracy 0.780 (AUROC 0.844). Accuracy drops on states longer than about 400 tokens; training data and benchmarks are both synthetic, from one generator model; the maker says not to use it for hiring, credit, medical or legal decisions. Needs the sokudan package (custom scorer and ordinal heads); sokudan serve offers a /v1/systemone-compatible server, and an MLX path runs on Apple Silicon. The maker did not measure TypeSafe Jev, citing Jev's terms.
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 probability
At a glance
| Parameters | 315M |
| Base model | sbintuitions/modernbert-ja-310m |
| Maker | GeneLab |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 88.0% |
| Reported latency | 22.8 ms p50 / 27.5 ms p95 per three-question request on an RTX 5090; 371 ms on a Core Ultra 9 285K CPU; about 10 ms per short question with MLX on an M1 Max |
Get the weights
pip install systemonemodels
systemone pull genelab/sokudan-ja-310m
The files are served from the maker's Hugging Face repository, GeneLab/sokudan-ja-310m, and verified against the checksums recorded here.