Zero-shot classification with Laya

Example

With the official package, following the core README:

python -m pip install laya

from laya import Router

router = Router()  # downloads a checkpoint on first use

questions = {
    "topic": {"type": "choice", "instructions": "What is this message about?",
              "criteria": {"billing": "invoices, payments, refunds",
                           "technical": "bugs, outages, system errors",
                           "account": "login, password, profile changes",
                           "other": "everything else"}},
    "mood": {"type": "score", "instructions": "How does the customer feel?",
             "criteria": ["happy", "neutral", "annoyed", "angry"]},
    "refund": {"type": "noul", "instructions": "Does the customer ask for a refund?"},
}

result = router.predict("I was charged twice for March and nobody answers my emails.", questions)
print(result["answers"]["topic"]["choice"])         # the most likely label
print(result["answers"]["topic"]["probabilities"])  # one probability per label
print(result["answers"]["refund"]["noul"])          # probability that the answer is yes

The labels live in the request, so changing the label set is a code change, not a training run.

Three kinds of question

• choice picks one label from a set: topic, intent, department. Describe each label in a few words, as in the README examples.
• score places the text on an ordered scale: sentiment, urgency, severity.
• noul returns the probability that a statement is true. For multi-label tagging, ask one noul question per tag.

Compared with NLI zero-shot pipelines

The classic approach, from Yin et al. (2019), reuses a model trained on natural language inference, such as facebook/bart-large-mnli (MIT licence). Each candidate label becomes a hypothesis like "This example is {label}." and the model scores whether the text entails it. The pipeline runs one text-hypothesis pair per label, so the cost grows with the number of labels. Laya scores all options of a question, and all questions in a call, in one forward pass.

For measured comparisons, jevbench (dhruvmehra) runs Jev, Laya, LLMs, a fine-tuned DistilBERT and zero-shot NLI on SST-2, AG News and Banking77. ollaya serves Laya next to zero-shot NLI and GLiClass classifiers behind one API, which makes a side-by-side test on your own data easy.

Compared with GLiNER-style models

GLiNER started as a zero-shot named-entity model: you pass entity types at inference time and it extracts matching spans. That is extraction, not classification. Fastino's GLiNER2 family added classification, and GLiNER2.5-Decide is a 340M English classifier that takes any label set at call time and supports multi-label output, under Apache 2.0. Its card reports 60.2% exact-match accuracy on Fastino's 17-domain fast-decisions suite.

In akitaonrails' benchmark, GLiNER2.5-Decide scored 52.4% on the compact set and Laya 42.9%, at 208 ms and 41 ms median latency on the author's machine. Both were far below the 4B+ decision models in the same test.

How accurate is zero-shot Laya?

It depends on the task:

• Few broad labels work well, subtle or many labels much less. The per-dataset numbers (AG News, DAIR Emotion, Banking77) are on Laya for text classification. Options share a token budget, so keep a question under about 20 options or split it into coarse and fine questions.
• Hard decisions sit near chance. The model card puts the base checkpoints at 0.362 and 0.342 on the typed-decisions benchmark, against 0.318 for random guessing and 0.461 for always picking the most common answer. Arize's benchmark found laya-multilingual at 59.8% balanced accuracy on RAGTruth hallucination detection, against 85.5% for Jev.
• Wording matters. In the same Arize test, flipping the question from "Is every claim supported?" to "Does it contain an unsupported claim?" moved Laya's ROC AUC from 0.64 to 0.36. The core README also warns that noul can follow its option labels instead of the text, and suggests a two-option choice with neutral keys as a workaround.

When zero-shot is not enough

Label a few hundred real examples and measure first. If the numbers are too low, you have two routes. Fine-tune: the laya-typed-decisions checkpoint went from 0.362 to 0.766 on its benchmark, and the notebook runs on Kaggle's free GPUs. Or move to a larger decision model; see Laya alternatives. More on labels and benchmarks is in Text classification with Laya and ModernBERT and mmBERT in Laya.