Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

nirnay

NIRNAY 450M: open Apache-2.0 decision model fine-tuned from Laya that reports 87.9% on Banking77 with fitted ECE 0.045.

ModelresearchApache-2.0
View on GitHub ↗Live demo ↗

Overview

NIRNAY (Hindi for decision) is a 450M open-weight classifier for banking intent and typed decisions from Eulogik. One forward pass turns a state plus a question into calibrated probabilities, with no text generation, API key or per-call bill. The list summary describes it as a Laya fork with a concept bottleneck, a coarse-to-fine pointer and RLCD.

On the Banking77 test set (3,080 cases), the README reports 0.8792 accuracy with Brier 0.208 and fitted ECE 0.045 for phase_b, against 0.803 for Jev 1.13.0 on the same cases. It notes the caveat that NIRNAY was fine-tuned on the train split while Jev answered zero-shot, arguing that this is the point of a model you fine-tune on your own data. Weights, a live demo space, code, a paper and raw eval JSON are published. It suits teams needing an on-device or self-hosted intent classifier.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich of the 77 banking intents does this message express?—

How it works

  1. Load the 450M model from Hugging Face.
  2. Send a customer message and the intent question.
  3. Get calibrated probabilities over intents.
  4. Act on high-confidence intents and escalate the rest.

Strengths

  • No API key or per-call cost.
  • Published raw evaluation JSON.
  • Calibration reported both raw and fitted.

Caveats

  • Comparison with Jev is fine-tuned versus zero-shot.
  • Specialised to banking intents.

Getting started

Try the Hugging Face demo space or download eulogik/nirnay-450m from Hugging Face.

At a glance

Type
Model
Status
research
Language
Python
Runtime
PyTorch
Has code
yes
License
Apache-2.0
Stars
9

Use cases and integrations

How curated lists describe it

  • NIRNAY 450M: open-weight calibrated decision model (Apache-2.0). Banking77 intent classification 87.9%, fitted ECE 0.045. Laya fork plus concept bottleneck, coarse-to-fine pointer and RLCD. Built by Eulogik.

    wh000wh000/awesome-jev-live

Frequently asked questions

Does NIRNAY really beat Jev?

On Banking77 it reports 87.9% vs 80.3%, but it was fine-tuned on that task while Jev was zero-shot.

What is it based on?

Per the list summary, a Laya fork with a concept bottleneck, coarse-to-fine pointer and RLCD.

Is it free to use?

Yes, weights are Apache-2.0 and run locally.

Related projects

  • sokudan

    Japanese System One decision model on ModernBERT-ja (314.6M) returning typed choice, score and bool distributions with no generation.

    ProjectbetaPython
  • systemANE

    Two-tier macOS decision engine: a MiniLM encoder on the Neural Engine decides, escalating unsure cases to Apple's on-device model.

    ProjectprototypePython
  • cbjev★ 4

    Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.

    ProjectresearchPython
  • bev

    Joke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.

    ProjectresearchPython
  • decision-model-audit

    Pre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.

    ProjectresearchPython
  • ej

    11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.

    ProjectresearchPython