Lavoir
A single-pass decision encoder on ModernBERT-large that returns a calibrated probability for every option and a value-of-information score for each candidate clarifying question, so it can ask, decide or hand off. Uses Laya's typed question format.
Follows Laya's decision-head design, input format and loss, but trains its own heads from ModernBERT-large and uses none of Laya's weights. A small value-of-information head estimates how much asking each missing slot would raise the probability of the right decision. Described in arXiv 2609.30706. Zero-shot on SGD, Mogan AI reports .942 accuracy while asking in 6% of conversations; on ABCD it reports .657 and is overconfident there (ECE .19). Workflows unseen in training are much weaker. A Turkish sibling, Lavoir-TR, is built on MoganBERT-TR. The weights are CC BY-NC 4.0 because some training data is non-commercial, so they are not mirrored or packaged here; the code is Apache-2.0.
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
- route — sends the input to one of several destinations
At a glance
| Base model | answerdotai/ModernBERT-large |
| Maker | Mogan AI |
| Released | 2026-09-25 |
| License | cc-by-nc-4.0 |
| Reported accuracy | 94.2% |
| Reported latency | 31 ms median, 142 ms p95 per question on one NVIDIA GH200 (bf16) |
Get the weights
pip install systemonemodels
systemone pull moganai/lavoir
The files are served from the maker's Hugging Face repository, moganai/lavoir, and verified against the checksums recorded here.
Read more
This page was opened by System One for Mogan AI, who can claim the organisation and take it over at any time.