Laya PT/ES Typed Decisions
Portuguese and Spanish fine-tune of Convai's multilingual Laya (322M, mmBERT). Answers choice, score and noul questions over a text state in one forward pass, with a temperature-scaled probability per option. Text only; Apache-2.0.
A full fine-tune with no LoRA: stage 1 on LocalLLaMA/typed-decisions translated to Portuguese and Spanish, stage 2 on 2,099 hard examples generated from Common Voice and MLS transcripts. Text only. Telepatia's own benchmark is 384 held-out translated typed-decisions questions, on which it scores 0.807 (Brier 0.111, ECE 0.027) against 0.362 for the base. Probabilities are temperature-scaled on the maker's validation set, so refit them on your own domain before using them as risk scores; score is the weakest primitive, and the benchmark covers four synthetic business workflows. Sibling: Hertz-1, which answers questions about Portuguese and Spanish audio through a speech recogniser feeding a frozen Laya checkpoint (a cascade with fixed option counts and mixed component licences), linked below rather than listed.
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 | 322M |
| Base model | convaiinnovations/laya-multilingual |
| Maker | Telepatia |
| Released | 2026-09-22 |
| License | apache-2.0 |
| Reported accuracy | 80.7% |
| Reported latency | 18.25 ms p50 / 18.69 ms p95, warm, batch size one, on a Modal A100-40GB |
Get the weights
pip install systemonemodels
systemone pull telepatia/laya-pt-es-typed
The files are served from the maker's Hugging Face repository, telepatia-ai/laya-pt-es-typed, and verified against the checksums recorded here.