Sciev v0.2.3
Typed scientific decision heads (choice, noul, score) that read a frozen LLaDA-8B-Instruct masked-diffusion backbone: one forward pass per decision, probabilities temperature-fitted on held-out dev data, no generated text. Heads only; the 8B backbone downloads separately.
An individual research release, not affiliated with TypeSafe AI. The heads are about 67M (choice) and 17M each (noul, score) on the frozen 8B backbone, which keeps its own MIT licence. The maker notes that licensing of the upstream training QA data is not fully established. The numbers come from the maker's own scientific battery with constructed reference labels; with the passage removed the choice head still matches the reference on 87.5% of items. Out of domain it is much weaker (GPQA main 0.295), and calibration fitted in-domain does not transfer (GPQA ECE 0.236), so refit the temperatures. pip install sciev; sciev[serve] runs a local /v1/systemone endpoint. The experimental heads in the repo need an adapter that was not released and are not listed.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one probability
- score — places the input on an ordered scale
At a glance
| Parameters | 8B |
| Base model | GSAI-ML/LLaDA-8B-Instruct |
| Maker | Angel Luis Robles-Fernández |
| Released | 2026-09-28 |
| License | apache-2.0 |
| Reported accuracy | 97.0% |
Get the weights
pip install systemonemodels
systemone pull alrobles/sciev
The files are served from the maker's Hugging Face repository, alrobles/sciev, and verified against the checksums recorded here.