BEDR-v2-Experiments-Data

Overview

This dataset supports the paper "BEDR: Boundary Entropy-Driven Resampling with Large Language Models for Fine-Grained Cyber Threat Intelligence Mapping". It contains the complete set of training and test data constructed by the BEDR framework for fine-grained Cyber Threat Intelligence (CTI) mapping to MITRE ATT&CK techniques (679 fine-grained classes).

The BEDR framework generates candidate sentences with a structured, knowledge-enhanced prompt, screens them by local boundary entropy computed in an intent-anchored space, raises long-tail classes through an adaptive diagnosis-refill loop, and compresses majority classes through entropy-guided selective undersampling.

Contents

The data/ directory contains the following files:

File	Description
D_bedrv2_HYBRID_top100.npz	Final BEDR label set: 24,174 samples, 679 classes, 30-100 samples per class (primary dataset)
D_bedrv2_HYBRID.npz	BEDR-v2 hybrid-arm dataset before head undersampling
D_bedrv2_PLAIN.npz	BEDR-v2 plain-arm dataset
D_bedrv2_NONE.npz	BEDR-v2 no-filter arm dataset
D_naive_full.npz	Naive LLM augmentation baseline: 34,153 samples, 679 classes, 40-80 per class
D_bedr_train.npz	Original BEDR training set
D_base_train_new.npz	Baseline training set from official ATT&CK procedure examples (11,840 samples, 574 classes)
test2_full_plain.npz	Real-sentence test set (2,984 real sentences, 420 classes; src marks real/generated)
orig4291_plain.npz	Full-coverage test set (4,291 samples, 679 classes)
classes_sorted.json	Ordered list of 679 technique categories
zero_shot_classes.json	The 105 zero-shot classes
intent_embeds.npz	Official intent embeddings per class
intent_action_entity.json	Structured attack graph G = {I, A, E} for 679 classes
enterprise-attack.json	MITRE ATT&CK v17.1 official knowledge base (STIX format)
filter_decision.npz	Plain-space local-entropy screening decisions
all_samples_intent_sentence_embed.npz	Intent-anchored embeddings of all samples
candidates/	Raw candidate sentences of rounds 5/6/6b/7 (CSV)
audit/	Embeddings and texts of the fidelity audit (removed/kept/naive vs official definitions)

Loading the primary dataset

import numpy as np
z = np.load('data/D_bedrv2_HYBRID_top100.npz', allow_pickle=True)
vectors = z['vectors']      # 768-dimensional plain sentence embeddings
labels = z['labels']        # class indices (0-678, ordered by classes_sorted.json)

Download

pip install huggingface_hub
huggingface-cli download Andou2yu/BEDR-v2-Experiments-Data --repo-type dataset --local-dir data

Related resources

• Experiment code, models, and results: Yu000910/BEDR-v2-Experiments

Citation

If you use this dataset, please cite:

@article{bedr2026,
  title={BEDR: Boundary Entropy-Driven Resampling with Large Language Models for Fine-Grained Cyber Threat Intelligence Mapping},
  author={Yu, Fengrui and Du, Yanhui},
  journal={Information Processing and Management},
  year={2026},
  note={under review}
}

License

MIT License. Note that enterprise-attack.json is from the MITRE ATT&CK framework and is subject to MITRE's licensing terms.