NIRNAY 450M
A 450M decision model from Eulogik: Laya plus a concept bottleneck and a coarse-to-fine pointer, fine-tuned on Banking77 intent. Answers choice, score and yes/no questions with a probability per option in one forward pass. Tuned for banking intent; measure before other uses.
A LoRA on a frozen Laya encoder plus about 30M new parameters, fine-tuned on the Banking77 train split; the maker calls English banking intent, triage and guardrails its only proven lane. On the Banking77 test split Eulogik reports accuracy 0.879, Brier 0.208 and macro-F1 0.879; raw ECE is 0.089, and the card's 0.045 "fitted" figure uses a temperature fitted on the test set itself. Eulogik's own run of the third-party JevBench public set scores 0.550 overall and 0.396 on the hard tier. 512-token context, longer inputs head-truncated. Loads only through the nirnay package from GitHub: phase_b.pt is a PyTorch pickle applied on top of the Laya base weights; an ONNX export is in the repo. The package includes a server with a Jev-compatible /v1/systemone endpoint.
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 | 450M |
| Base model | convai-innovations/laya |
| Maker | Eulogik |
| Released | 2026-09-30 |
| License | apache-2.0 |
| Reported accuracy | 87.9% |
| Reported latency | 209 ms per decision on an Apple M4 GPU (MPS), 361 ms on CPU, batch 1 |
Get the weights
pip install systemonemodels
systemone pull eulogik/nirnay
The files are served from the maker's Hugging Face repository, eulogik/nirnay-450m, and verified against the checksums recorded here.
Read more
This page was opened by System One for Eulogik, who can claim the organisation and take it over at any time.