README.md

---
license: cc0-1.0
configs:
- config_name: default
  data_files:
  - split: wikidata
    path: ReFactX_wikidata_facts.bz2
task_categories:
- question-answering
---

# ReFactX: Scalable Reasoning with Reliable Facts via Constrained Generation Dataset

This dataset contains verbalized facts from Wikidata to simplify reproducing [ReFactX](https://doi.org/10.1007/978-3-032-09527-5_16). ReFactX is a scalable method that enables Large Language Models (LLMs) to access external knowledge without depending on retrievers or auxiliary models. It utilizes constrained generation with a pre-built prefix-tree index, verbalizing triples from a Knowledge Graph into textual facts, tokenizing them, and indexing them for efficient access. The dataset specifically includes `ReFactX_wikidata_facts.bz2`, which consists of 800 million textual facts from Wikidata, designed for use in the ReFactX system.

## Paper

[The Semantic Web - ISWC 2025](https://doi.org/10.1007/978-3-032-09527-5_16)

[Hugging Face Paper Page](https://huggingface.co/papers/2508.16983)

[Arxiv Preprint](https://arxiv.org/abs/2508.16983)

## Code

[GitHub Repository](https://github.com/rpo19/ReFactX)

## Abstract

Knowledge gaps and hallucinations are persistent challenges for Large Language Models (LLMs), which generate unreliable responses when lacking the necessary information to fulfill user instructions. Existing approaches, such as Retrieval-Augmented Generation (RAG) and tool use, aim to address these issues by incorporating external knowledge. Yet, they rely on additional models or services, resulting in complex pipelines, potential error propagation, and often requiring the model to process a large number of tokens. In this paper, we present a scalable method that enables LLMs to access external knowledge without depending on retrievers or auxiliary models. Our approach uses constrained generation with a pre-built prefix-tree index. Triples from a Knowledge Graph are verbalized in textual facts, tokenized, and indexed in a prefix tree for efficient access. During inference, to acquire external knowledge, the LLM generates facts with constrained generation which allows only sequences of tokens that form an existing fact. We evaluate our proposal on Question Answering and show that it scales to large knowledge bases (800 million facts), adapts to domain-specific data, and achieves effective results. These gains come with minimal generation-time overhead.

## Sample Usage

To get started with ReFactX, first install the required dependencies:

```bash
pip install -r requirements.txt
```

For quickly trying ReFactX with an in-memory prefix tree (derived from a 31k-facts knowledge base), use the notebook [`try_refactx.ipynb`](https://github.com/rpo19/ReFactX/blob/main/try_refactx.ipynb) from the [GitHub repository](https://github.com/rpo19/ReFactX).

To reproduce the experiments described in the paper, refer to the `eval.py` script in the GitHub repository, replacing `INDEX`, `MODEL`, and `DATASET` according to your specific needs.

## Citation

```bibtex
@InProceedings{10.1007/978-3-032-09527-5_16,
      author="Pozzi, Riccardo
      and Palmonari, Matteo
      and Coletta, Andrea
      and Bellomarini, Luigi
      and Lehmann, Jens
      and Vahdati, Sahar",
      editor="Garijo, Daniel
      and Kirrane, Sabrina
      and Salatino, Angelo
      and Shimizu, Cogan
      and Acosta, Maribel
      and Nuzzolese, Andrea Giovanni
      and Ferrada, Sebasti{\'a}n
      and Soulard, Thibaut
      and Kozaki, Kouji
      and Takeda, Hideaki
      and Gentile, Anna Lisa",
      title="ReFactX: Scalable Reasoning with Reliable Facts via Constrained Generation",
      booktitle="The Semantic Web -- ISWC 2025",
      year="2026",
      publisher="Springer Nature Switzerland",
      address="Cham",
      pages="290--308",
      isbn="978-3-032-09527-5",
      doi="10.1007/978-3-032-09527-5_16",
      url="https://doi.org/10.1007/978-3-032-09527-5_16"
}
```