UltraGliClass v0.1

UltraGliClass v0.1

A 154M-parameter zero-shot text classifier that scores every label you pass in one pass, for single-label, multi-label and intent tasks, including long inputs. For developers who need quick text labeling in Python.

Made by Srimon12 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: UltraGliClass v0.1 labels text against categories you name on the spot, using a compact model of roughly 150M parameters. You pass a text, a list of labels and, if you like, an instruction. It scores every label in one forward pass, so you do not need a training set for your own labels. It uses the GLiClass head on top of the UltraBERT-150M encoder.

Who it suits: Developers and data teams who need to label text by categories they define themselves, including cases with many labels or long inputs.

Industry | Who uses it | How they use it | Example
Customer support | Teams routing messages | Match a message to one of many intents | Label a banking request among dozens of intents
Data engineering | Pipeline builders | Tag documents with custom labels | Mark a long record as one of several decisions
Research | NLP researchers | Compare zero-shot classifiers on long inputs | Test choice and yes-no decisions at 8,192 tokens

You can skip it if you need a model for languages other than English, since the card lists English only.

Where to use it:

• Python with the transformers and gliclass packages
• Windows, macOS or Linux, wherever Python runs
• Text classification steps inside your own pipelines

Use case | Who | What they get back
Single-label classification | Developers | One best label for a text
Multi-label classification | Data teams | A score for every label you listed
Intent detection with many labels | Support teams | Scores across dozens or hundreds of intents
Long decisions | Analysts | Choice, yes-no or score calls over long inputs

How to set it up: You need Python and the two packages at the versions the card used for training and evaluation. Transformers 5.20 and later supports the encoder natively, so no trust_remote_code is needed.

• Python and pip
• transformers 5.20.2 or newer
• gliclass 0.1.21

1. Install the two packages.
2. Download the model, or let the library fetch it by name.
3. Load the model and tokenizer in Python.

pip install "transformers>=5.20.2" "gliclass==0.1.21"

pip install -U huggingface_hub
hf download Srimon12/ultragliclass-v0.1

How to use it: Load the model and the tokenizer, then build a zero-shot classification pipeline. The pipeline takes one text, your candidate labels as strings, and an instruction if you want one.

import torch
from gliclass import GLiClassModel, ZeroShotClassificationPipeline
from transformers import AutoTokenizer

repo = "Srimon12/ultragliclass-v0.1"
model = GLiClassModel.from_pretrained(repo)
tokenizer = AutoTokenizer.from_pretrained(repo, add_prefix_space=False)  # trained without a prefix space

Tool | What it does
GLiClassModel | Loads the classifier weights
AutoTokenizer | Loads the tokenizer, with add_prefix_space set to False
ZeroShotClassificationPipeline | Scores each label for a text in one pass

Example: You pass the text "My card was stolen, please stop it" with the labels lost card, balance enquiry and transfer. The pipeline returns a score for each label, and you take the highest one as the intent.

For newcomers and experts: If you are new, think of it as sorting text into boxes you name yourself. Start with two or three short labels and a few example sentences before you try a long list.

For experts, training ran in three stages. The last stage used inputs as long as 8,192 tokens and as many as 256 labels. The encoder supports 32,768 positions, but nothing beyond 8,192 tokens was tested. The trade-off is that it trails gliclass-base-v3.0 on a few sets such as sst5 and 20 Newsgroups.

Text, labels and optional instruction
UltraBERT encoder with GLiClass head
One forward pass
A score for every label

Good to know:

Title: Average macro-F1 on the 14-set zero-shot suite
UltraGliClass v0.1: 0.688
gliclass-base-v3.0: 0.673
gliclass-modern-base-v3.0: 0.558

• Language: English
• Length: trained and tested at 8,192 tokens, and longer inputs are untested
• Results: the figures come from the author's own run, so check them on your data
• License: Apache 2.0, with a notice on the card that some training rows carry a share-alike license, so read it before commercial use
• Cost to run: no API key or paid service named
• Data it reaches: downloads files from Hugging Face, and your text stays in your own Python process
• Not for: languages other than English, or inputs past 8,192 tokens without your own testing

On Hugging Face, 13 November 2026

• Task: zero shot classification
• Library: gliclass
• Parameters: 154M
• Weights: about 615 MB in F32; about 77 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 59
• Likes: 4
• License: apache-2.0
• Languages: en
• Based on: dmis-lab/UltraBERT-150M
• Tags: gliclass, safetensors, zero-shot-classification, text-classification, long-context, decision-model, ultrabert

Links

• Model: https://huggingface.co/Srimon12/ultragliclass-v0.1
• Files: https://huggingface.co/Srimon12/ultragliclass-v0.1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Srimon12/ultragliclass-v0.1

Then follow the usage in its model card. Nothing in this kit runs by itself.