Hugging Face's logo

knowledgator
/
gliclass-base-v2.0-rac-init

Zero-Shot Classification
Safetensors
English
French
ge
GLiClass
text classification
zero-shot
small language models
RAG
sentiment analysis
Model card Files Community
gliclass-base-v2.0-rac-init / README.md
BioMike's picture
BioMike
Update README.md
4611dd4 verified
preview code
|
raw
history blame
10.2 kB
metadata
license: apache-2.0
datasets:
  - MoritzLaurer/synthetic_zeroshot_mixtral_v0.1
  - knowledgator/gliclass-v1.0
  - fancyzhx/amazon_polarity
  - cnmoro/QuestionClassification
  - Arsive/toxicity_classification_jigsaw
  - shishir-dwi/News-Article-Categorization_IAB
  - SetFit/qnli
  - nyu-mll/multi_nli
  - SetFit/student-question-categories
  - SetFit/tweet_sentiment_extraction
  - SetFit/hate_speech18
  - saattrupdan/doc-nli
  - knowledgator/gliclass-v2.0-RAC
language:
  - en
  - fr
  - ge
metrics:
  - f1
pipeline_tag: zero-shot-classification
tags:
  - text classification
  - zero-shot
  - small language models
  - RAG
  - sentiment analysis
base_model:
  - microsoft/deberta-v3-base

⭐ GLiClass: Generalist and Lightweight Model for Sequence Classification

This is an efficient zero-shot classifier inspired by GLiNER work. It demonstrates the same performance as a cross-encoder while being more compute-efficient because classification is done at a single forward path.

It can be used for topic classification, sentiment analysis and as a reranker in RAG pipelines.

The model was trained on synthetic and licensed data that allow commercial use and can be used in commercial applications.

This version of the model uses a layer-wise selection of features that enables a better understanding of different levels of language. The backbone model is ModernBERT-base, which effectively processes long sequences.

Retrieval-augmented Classification (RAC):

The main idea of this model is to utilize the information from semantically similar examples to enhance predictions in inference. The tests showed that providing the model with at least one example from the train dataset, which was retrieved by semantic similarity, could increase the F1 score from 0.3090 to 0.4275, in some cases from 0.2594 up to 0.6249. Moreover, the RAC approach, with 2 examples provided, showed an F1 score, compared to fine-tuning with 8 examples per label: 0.4707 and 0.4838, respectively.

RAC dataset generation strategy:

image/png image/png To further enhance classification performance, we generated a Retrieval-Augmented Classification (RAC) dataset. Each text example in the gliclass-v2.0 dataset was encoded using the paraphrase-MiniLM-L6-v2 sentence transformer and indexed in an HNSW (Hierarchical Navigable Small World) database. For 250k randomly selected samples, we retrieved up to three most similar examples (cosine similarity > 0.5) from the dataset.

During augmentation:

  • The number of retrieved examples per sample was randomly chosen between 1 and 3.
  • 30% of retrieved examples were replaced with random, unrelated examples to introduce controlled noise.
  • If true labels were present in a retrieved example, false labels were removed with a 50% probability to balance information clarity.

Each retrieved example was formatted using structured <<EXAMPLE>> ... <</EXAMPLE>> tags, where:

  • True labels were explicitly marked as <<TRUE_LABEL>> {label}.
  • False labels were marked as <<FALSE_LABEL>> {label}, unless removed.

For each randomly selected 250k examples, the “text” was modified as {original_text} <<EXAMPLE>> {retrieved_text} {true_labels_str} {false_labels_str} <</EXAMPLE>>... Where:

  • {original_text} is the original example text.
  • {retrieved_text} is a similar or randomly selected example.
  • {true_labels_str} contains true labels formatted as <<TRUE_LABEL>> {label}.
  • {false_labels_str} contains false labels formatted as <<FALSE_LABEL>> {label} (unless removed with 50% probability).

Such a strategy allows the model to learn how to utilize the provided information without overfocusing on RAC examples. With both relevant and randomly retrieved examples, the dataset maintains a balance between useful contextual information and controlled noise. This ensures that the model does not become overly reliant on retrieval-augmented inputs while still benefiting from additional context when available.

How to use:

First of all, you need to install GLiClass library:

pip install gliclass

Than you need to initialize a model and a pipeline:

from gliclass import GLiClassModel, ZeroShotClassificationPipeline
from transformers import AutoTokenizer

model = GLiClassModel.from_pretrained("knowledgator/gliclass-modern-base-v2.0-init")
tokenizer = AutoTokenizer.from_pretrained("knowledgator/gliclass-modern-base-v2.0-init")
pipeline = ZeroShotClassificationPipeline(model, tokenizer, classification_type='multi-label', device='cuda:0')

text = "One day I will see the world!"
labels = ["travel", "dreams", "sport", "science", "politics"]
results = pipeline(text, labels, threshold=0.5)[0] #because we have one text
for result in results:
 print(result["label"], "=>", result["score"])

If you want to use it for NLI type of tasks, we recommend representing your premise as a text and hypothesis as a label, you can put several hypotheses, but the model works best with a single input hypothesis.

# Initialize model and multi-label pipeline
text = "The cat slept on the windowsill all afternoon"
labels = ["The cat was awake and playing outside."]
results = pipeline(text, labels, threshold=0.0)[0]
print(results)

Benchmarks:

Below, you can see the F1 score on several text classification datasets. All tested models were not fine-tuned on those datasets and were tested in a zero-shot setting.

Model IMDB AG_NEWS Emotions
gliclass-modern-large-v2.0-init (399 M) 0.9137 0.7357 0.4140
gliclass-modern-base-v2.0-init (151 M) 0.8264 0.6637 0.2985
gliclass-large-v1.0 (438 M) 0.9404 0.7516 0.4874
gliclass-base-v1.0 (186 M) 0.8650 0.6837 0.4749
gliclass-small-v1.0 (144 M) 0.8650 0.6805 0.4664
Bart-large-mnli (407 M) 0.89 0.6887 0.3765
Deberta-base-v3 (184 M) 0.85 0.6455 0.5095
Comprehendo (184M) 0.90 0.7982 0.5660
SetFit BAAI/bge-small-en-v1.5 (33.4M) 0.86 0.5636 0.5754

Below you can find a comparison with other GLiClass models:

Dataset gliclass-base-v1.0-init gliclass-large-v1.0-init gliclass-modern-base-v2.0-init gliclass-modern-large-v2.0-init
CR 0.8672 0.8024 0.9041 0.8980
sst2 0.8342 0.8734 0.9011 0.9434
sst5 0.2048 0.1638 0.1972 0.1123
20_news_groups 0.2317 0.4151 0.2448 0.2792
spam 0.5963 0.5407 0.5074 0.6364
financial_phrasebank 0.3594 0.3705 0.2537 0.2562
imdb 0.8772 0.8836 0.8255 0.9137
ag_news 0.5614 0.7069 0.6050 0.6933
emotion 0.2865 0.3840 0.2474 0.3746
cap_sotu 0.3966 0.4353 0.2929 0.2919
rotten_tomatoes 0.6626 0.7933 0.6630 0.5928
AVERAGE: 0.5344 0.5790 0.5129 0.5447

Here you can see how the performance of the model grows providing more examples:

Model Num Examples sst5 ag_news emotion AVERAGE:
gliclass-modern-large-v2.0-init 0 0.1123 0.6933 0.3746 0.3934
gliclass-modern-large-v2.0-init 8 0.5098 0.8339 0.5010 0.6149
gliclass-modern-large-v2.0-init Weak Supervision 0.0951 0.6478 0.4520 0.3983
gliclass-modern-base-v2.0-init 0 0.1972 0.6050 0.2474 0.3499
gliclass-modern-base-v2.0-init 8 0.3604 0.7481 0.4420 0.5168
gliclass-modern-base-v2.0-init Weak Supervision 0.1599 0.5713 0.3216 0.3509
gliclass-large-v1.0-init 0 0.1639 0.7069 0.3840 0.4183
gliclass-large-v1.0-init 8 0.4226 0.8415 0.4886 0.5842
gliclass-large-v1.0-init Weak Supervision 0.1689 0.7051 0.4586 0.4442
gliclass-base-v1.0-init 0 0.2048 0.5614 0.2865 0.3509
gliclass-base-v1.0-init 8 0.2007 0.8359 0.4856 0.5074
gliclass-base-v1.0-init Weak Supervision 0.0681 0.6627 0.3066 0.3458