ENPT

OpenDecision

Last updated:

WrapperApache-2.0~57 ★

⚠️ Why not a model

Wraps a stock, frozen zero-shot NLI classifier (ModernBERT-large) behind a Jev-style API. No decision-specific training; the author says the scores are uncalibrated.

Technical specs

Base LLMs
MoritzLaurer/ModernBERT-large-zeroshot-v2.0 (default), MoritzLaurer/deberta-v3-large-zeroshot-v2.0
Decision types
choice, noul, score
Features
local, systemone-endpoint, typesafe-sdk-compatible, zero-shot-nli, document-api, evidence-retrieval, relation-primitive, pypi
License
Apache-2.0

Links

Not to be confused with OpenDecisions (the OneJev server) or the OpenDecider model.

OpenDecision is an Apache-2.0 project by deepanwadhwa (created 17/09/2026, PyPI OpenDecision 0.1.2) that calls itself “the open-source equivalent of TypeSafe’s Jev”. It accepts state or a document plus typed questions and returns structured answers without generating text.

How it works

  • Backend: the off-the-shelf MoritzLaurer/ModernBERT-large-zeroshot-v2.0 NLI classifier (~400M parameters), or optionally deberta-v3-large-zeroshot-v2.0. Each option is scored as an entailment hypothesis. No weights are trained for decisions.
  • Primitives: Choice, Noul, Score, plus Relation (supports / contradicts / unknown / conflicted).
  • Documents: POST /v1/documents/decide splits long documents, retrieves passages and returns the ones it used; Noul has binary, three_way and both modes.
  • TypeSafe compatibility: POST /v1/systemone accepts the core TypeSafe request shape, so a local server can be the base_url for compatible clients.
  • Demo: picks actions for a ViZDoom bot in real time (fixed-seed recordings; no win rate measured).

Why it is here and not in the catalog

It was queued for a catalog evaluation since 22/09. The decisive facts: the model is a stock zero-shot classifier with no decision training, and the README says to “treat the model scores as uncalibrated”. That is a wrapper over an existing model, the same reasoning used for simple-jev and SemIf. It would move to the catalog if a fine-tuned, calibrated checkpoint is published.

Status

Developer preview (v0.1.2), Python 3.13+. 57 stars, 3 forks, last push 25/09 (01/10 check). Results shown (insurance claim case, GDPR 9 of 10) are single development cases.