Malkuth-4B
Multilingual decision model with a Korean focus: a LoRA adapter and pointer head post-trained from Kev on frozen Qwen3.5-4B-Base, returning probabilities for choice, noul and score questions. Research-use-only weights.
A personal release by Saechan Oh. The repo holds the adapter only (33.8M trainable parameters plus head.pt, a pickle); the base downloads separately. Served with Kev's server, so a TypeSafe-style client works with a base URL change. Initialised from jaredpalmer/kev-4b; the published weights are an intermediate checkpoint chosen on the validation split. On the maker's own 29-suite evaluation, held-out accuracy (9 suites not trained on) is 0.724; 10 of the 29 suites, including 4 of 5 Korean ones, were in training, and the maker's figures show the probabilities are overconfident. The weights are research use only (XNLI and RACE in the training data); the code is Apache-2.0. Sibling Malkuth-2B.
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
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 4B |
| Base model | qwen/qwen3.5-4b-base |
| Maker | Saechan Oh |
| Released | 2026-09-24 |
| License | cc-by-nc-4.0 |
| Reported accuracy | 72.4% |
Get the weights
pip install systemonemodels
systemone pull saechan-oh/malkuth
The files are served from the maker's Hugging Face repository, dhtocks/malkuth-4b, and verified against the checksums recorded here.
Read more
This page was opened by System One for Saechan Oh, who can claim the organisation and take it over at any time.