metadata
language:
- en
license: apache-2.0
library_name: transformers
tags:
- text-classification
- hallucination-detection
- grounding
- factual-consistency
- nli
- rag
datasets:
- stanfordnlp/snli
- nyu-mll/multi_nli
- anli
pipeline_tag: text-classification
🛡️ FactGuard
Lightweight hallucination and grounding detection model. Checks whether a claim is supported by the given context.
Built on ModernBERT-base (149M params), fine-tuned on 1M+ NLI pairs from SNLI, MultiNLI, and ANLI.
Classes: Supported, Not Supported
🚀 Usage
from transformers import pipeline
classifier = pipeline("text-classification", model="ENTUM-AI/FactGuard")
result = classifier({
"text": "Apple reported revenue of $94.8 billion in Q1 2024.",
"text_pair": "Apple's Q1 2024 revenue was $94.8 billion."
})
# [{'label': 'Supported', 'score': 0.99}]
result = classifier({
"text": "Apple reported revenue of $94.8 billion in Q1 2024.",
"text_pair": "Apple's revenue exceeded $100 billion."
})
# [{'label': 'Not Supported', 'score': 0.97}]
📊 Training Data
| Dataset | Samples |
|---|---|
| stanfordnlp/snli | ~550K |
| nyu-mll/multi_nli | ~393K |
| anli | ~163K |
1M+ NLI pairs mapped to binary grounding labels.
🔍 Use Cases
- RAG pipelines — verify LLM responses against source documents
- Fact-checking — detect unsupported claims in generated text
- Content moderation — flag hallucinated content before publishing
⚠️ Limitations
- English only
- Designed for single claim verification against a given context