JevK5-Lite
A 437M DeBERTa-v3-large CPU classifier from the JevK5 project. It reads a text and any number of caller-supplied label sets in one encoder pass and returns a temperature-scaled probability per label: softmax within a single-label head, sigmoid per label in a multi-label head.
Released as preview-1 and labelled experimental by its author; English only. On fastino/fast-decisions dev the maker reports head accuracy 0.587 against GLiNER2.5-Decide's 0.637. On seven public datasets fixed in advance, its mean macro-F1 is 0.629 against 0.643: it wins on AG News and Yahoo Answers and loses on Financial PhraseBank and GoEmotions. Its calibration error is lower on six single-label sets (0.035 to 0.103), but not under every shift (ECE 0.255 on fast-decisions dev). The author's own run on JevBench's 231 public items, a third-party suite, through an adapter that maps choice, score and noul onto label heads, scores easy 0.958, standard 0.750 and hard 0.405. Loads through the jevk5 package's JevK5Lite class. 20,822 training documents were written by GPT-6 Luna and are subject to OpenAI's terms.
What it decides
- classify — assigns a category from a fixed taxonomy
- choice — picks one option from a set
At a glance
| Parameters | 437M |
| Base model | microsoft/deberta-v3-large |
| Maker | Alibi Serikbay |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 58.7% |
| Reported latency | 54 to 92 ms median per item, bf16 on 16 CPU threads |
Get the weights
pip install systemonemodels
systemone pull alibi-serikbay/jevk5-lite
The files are served from the maker's Hugging Face repository, alibiserikbay/JevK5-Lite, and verified against the checksums recorded here.