NLP Zero-Shot Classify

Runs an ONNX Natural Language Inference (NLI) model to classify text against an arbitrary list of labels chosen at runtime. Unlike regular classification, the model is never trained on the specific labels - it's evaluating whether the input text entails each candidate label.

Slower than dedicated classifiers but flexible: change the labels per flow without retraining.

Field	Type	Default	Description
Name	string	required	Identifier for this pipeline instance. Must be unique within the flow
Path	string	required	Local path where the model lives. See Path and download modes below
Enable Download	boolean	false	Fetch the model from HuggingFace into Path at startup
Repository	string	-	HuggingFace repository to download from. Required when Enable Download is on
ONNX File Path	string	model.onnx	Path of the .onnx file inside the HuggingFace repository
Labels	list of strings	required	Candidate labels to classify against. At least one
Multi-Label	boolean	false	Score each label independently. Off means scores sum to 1 across labels
Hypothesis Template	string	This example is {}.	Template used to turn each label into an NLI hypothesis. Must contain {} where the label is inserted

Path and download modes

The Path field changes meaning based on Enable Download:

• Download off - Path points to an existing ONNX model on disk. Either a .onnx file directly, or a directory that contains model.onnx plus tokenizer.json. Qaynaq does not fetch anything; if the files are missing the flow fails at startup.
• Download on - Path is the parent directory where qaynaq will place downloaded models. Qaynaq creates a sub-directory inside it per repository (e.g. ./models/sentence-transformers_all-MiniLM-L6-v2/). The parent directory is created if missing. The download runs once; subsequent restarts reuse what's already on disk.

Sharing models across flows

Different flows that use the same repository can safely share the same Path - they both land in the same per-repository sub-directory and reuse the files. Different repositories under the same Path get their own sub-directories. ./models is a fine default for most setups.

A worker also keeps a content-addressed cache in ~/.cache/huggingface/hub/, so even unrelated Paths never re-download the same model from HuggingFace.

Output

{
  "sequence": "I am going to the park",
  "labels": ["fun", "boring", "dangerous"],
  "scores": [0.77, 0.15, 0.08]
}

Labels are returned sorted by score, highest first.

Hypothesis template

The template wraps each candidate label into a sentence the NLI model can score. The default "This example is {}." works for adjective-style labels like positive or boring. For other shapes:

Label style	Suggested template
Adjectives (busy, relaxed)	This person is {}.
Topics (sports, politics)	This text is about {}.
Intents (refund, complaint)	The user wants to {}.

Suggested models

Model	Notes
KnightsAnalytics/deberta-v3-base-zeroshot-v1	Strong general-purpose zero-shot
MoritzLaurer/mDeBERTa-v3-base-mnli-xnli	Multilingual