KnowRL-Knowledge-Base

Knowledge Base for "KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality"

  📄arXiv •   💻GitHub Repo •   🤗Models • 📚Training Data

Overview

This repository contains the external knowledge base used in the research paper, KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality.

The KnowRL framework is designed to address the issue of hallucinations in Large Language Models (LLMs), particularly in "slow-thinking" models that perform complex reasoning. It achieves this by integrating a factuality reward into the Reinforcement Learning (RL) process, guiding the model to ground its reasoning in verifiable facts. This knowledge base is the source of truth against which the model's statements are verified during training.

Dataset Description

This dataset is a crucial component for Stage: Knowledgeable Reinforcement Learning (RL). It serves as the external knowledge source used to calculate the Fact Reward, which evaluates the factuality of the model's reasoning process.

• Purpose: Provides the grounding knowledge for factuality verification during the RL training stage of the KnowRL framework.

• Data Source: The knowledge base is constructed from a snapshot of English Wikipedia (en.wikipedia.org) from November 1, 2023, containing 6.4 million entries.

• Data Format: Each entry consists of an entity (title) and the full text content from the corresponding Wikipedia page.

  {
    "title": "Specific Entity Name",
    "text": "Full text content from the corresponding Wikipedia page..."
  }
  

How to Use

Using the datasets Library

You can easily load the knowledge base using the datasets library in Python.

from datasets import load_dataset

# Load the knowledge base
knowledge_base = load_dataset("zjunlp/KnowRL-Knowledge-Base")

Using huggingface-cli

You can also download the entire dataset from the command line using huggingface-cli.

huggingface-cli download zjunlp/KnowRL-Knowledge-Base --repo-type dataset --local-dir KnowRL-Knowledge-Base

Dataset Construction Process

The knowledge base was constructed as part of the data preparation for the KnowRL training process.

1. Data Source: A comprehensive dump of English Wikipedia from November 1, 2023 was used as the foundation.

2. Entity Linking: During the RL data construction phase, questions were processed to extract key entities.

3. Knowledge Retrieval: For each question in the training data, the extracted entities were used to retrieve corresponding entries from the Wikipedia dump. An exact match or full containment of the entity within a Wikipedia entry title was required for a successful link.

4. Base Construction: The textual content of these successfully matched Wikipedia entries was compiled to form this knowledge base. To avoid redundancy, each entity keyword is linked to a maximum of three Wikipedia entries.

Citation

If you find this work useful in your research, please consider citing our paper:

@article{ren2025knowrl,
  title={{KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality}}, 
  author={Ren, Baochang and Qiao, Shuofei and Yu, Wenhao and Chen, Huajun and Zhang, Ningyu},
  journal={arXiv preprint arXiv:2506.19807},
  year={2025}

}