Verified, then logged.
Sun 22 Nov 2026 · 00:33 KST / 15:33 UTC한국어 | ENSearch /

02 / IT · Analysis

Hugging Face proposes a shared decision-model interface

The proposed interface brings probabilities for predefined answers and different question types under a common input and output contract.

Image from Hugging Face
Image: Hugging Face

Hugging Face announced decision-model support in its Transformers software library, without naming a licence in the announcement.

Its official post describes models that return probabilities for predefined answers instead of generating text. This lets applications ask for decisions within a set of answers they already recognise.

Hugging Face proposes a common input and output interface that accommodates different internal implementations. The aim is to give surrounding tools a shared contract for handling decisions across models.

Returning probabilities for predefined answers

The official post describes requests built from a state and questions. The state supplies the information to evaluate, as text or structured data; the questions specify what to assess and which answers are allowed.

The interface offers different answer types, according to the post. Choice selects an option and returns probabilities across the alternatives, while Score returns a score calculated by multiplying each ordered level by its probability and adding the results. Noul returns the probability that the answer to a yes-or-no question is true.

The post also describes a confidence value in Choice and Score responses, summarising how concentrated the probability distribution is. It distinguishes this value from the probability assigned to the most likely answer.

A common contract across different models

The official post compares decision models with zero-shot classification, which accepts candidate categories when a model runs. It presents the distinction as combining several questions and answer types in a consistent interface, rather than introducing probability-based outputs themselves.

Hugging Face's support announcement includes an example that loads autotrust/GEV-26B-Decide through Transformers. The post does not specify the example model's minimum VRAM requirement.

Hugging Face proposes extending the common contract from inference engines, which run models, to training tools. Models can retain different internal designs; consistent inputs and outputs would give those tools a shared basis for supporting them. The proposal centres on that agreement across the surrounding software.