KRLabsOrg/lettucedect-210m-eurobert-it-v1

LettuceDetect: Italian Hallucination Detection Model

Model Name: KRLabsOrg/lettucedect-210m-eurobert-it-v1 Organization: KRLabsOrg
Github: https://github.com/KRLabsOrg/LettuceDetect

Overview

LettuceDetect is a transformer-based model for hallucination detection on context and answer pairs, designed for multilingual Retrieval-Augmented Generation (RAG) applications. This model is built on EuroBERT-210M, which has been specifically chosen for its extended context support (up to 8192 tokens) and strong multilingual capabilities. This long-context capability is critical for tasks where detailed and extensive documents need to be processed to accurately determine if an answer is supported by the provided context.

This is our Italian base model utilizing EuroBERT-210M architecture

Model Details

• Architecture: EuroBERT-210M with extended context support (up to 8192 tokens)
• Task: Token Classification / Hallucination Detection
• Training Dataset: RagTruth-IT (translated from the original RAGTruth dataset)
• Language: Italian

How It Works

The model is trained to identify tokens in the Italian answer text that are not supported by the given context. During inference, the model returns token-level predictions which are then aggregated into spans. This allows users to see exactly which parts of the answer are considered hallucinated.

Usage

Installation

Install the 'lettucedetect' repository

pip install lettucedetect

Using the model

from lettucedetect.models.inference import HallucinationDetector

# For a transformer-based approach:
detector = HallucinationDetector(
    method="transformer", 
    model_path="KRLabsOrg/lettucedect-210m-eurobert-it-v1",
    lang="it",
    trust_remote_code=True
)

contexts = ["La Francia è un paese in Europa. La capitale della Francia è Parigi. La popolazione della Francia è di 67 milioni."]
question = "Qual è la capitale della Francia? Qual è la popolazione della Francia?"
answer = "La capitale della Francia è Parigi. La popolazione della Francia è di 69 milioni."

# Get span-level predictions indicating which parts of the answer are considered hallucinated.
predictions = detector.predict(context=contexts, question=question, answer=answer, output_format="spans")
print("Previsioni:", predictions)

# Previsioni: [{'start': 37, 'end': 83, 'confidence': 0.9231457829475403, 'text': ' La popolazione della Francia è di 69 milioni.'}]

Performance

Results on Translated RAGTruth-IT

We evaluate our Italian models on translated versions of the RAGTruth dataset. The EuroBERT-210M Italian model achieves an F1 score of 65.93%, outperforming prompt-based methods like GPT-4.1-mini (61.06%) with an improvement of +4.87 percentage points.

For detailed performance metrics, see the table below:

Language	Model	Precision (%)	Recall (%)	F1 (%)	GPT-4.1-mini F1 (%)	Δ F1 (%)
Italian	EuroBERT-210M	60.57	72.32	65.93	61.06	+4.87
Italian	EuroBERT-610M	76.67	72.85	74.71	61.06	+13.65

While the 610M variant achieves higher performance, the 210M model offers a good balance between accuracy and computational efficiency, processing examples approximately 3× faster. It shows particularly strong recall performance at 72.32%.

Citing

If you use the model or the tool, please cite the following paper:

@misc{Kovacs:2025,
      title={LettuceDetect: A Hallucination Detection Framework for RAG Applications}, 
      author={Ádám Kovács and Gábor Recski},
      year={2025},
      eprint={2502.17125},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.17125}, 
}