FSD-LAB/noc-incident-dataset-for-dl

NOC Incident Dataset for ART

This dataset contains preprocessed log-template sequences for training and evaluating ART (Abstract Reasoning LogTransformer) on network device incident prediction. Raw logs have already been parsed, anonymized, mapped to concept tokens, and grouped into model-ready sequence samples.

Data Format

The dataset is stored as daily Python pickle files:

• Files: YYYY-MM-DD.pkl
• Date range: 2025-11-21 to 2025-12-30
• Object type: list[tuple]

Each tuple follows this structure:

(concept_tokens, timestamp, device_id, label, concept_templates)

Field meanings:

• concept_tokens: list of integer concept-token IDs used by ART.
• timestamp: sample timestamp in UTC+7.
• device_id: integer device identifier.
• label: binary incident label for the sample.
• concept_templates: decrypted, human-readable log templates aligned with concept_tokens.

The final concept_templates element is included only for inspection and interpretability. Remove it before training or inference.

Loading Example

import pickle

with open("2025-11-21.pkl", "rb") as f:
    samples = pickle.load(f)

concept_tokens, timestamp, device_id, label, concept_templates = samples[0]

# Model input for ART: drop the decrypted concept-template field.
model_sample = (concept_tokens, timestamp, device_id, label)

Intended Use

This release is intended for sequence-based deep learning models for short-horizon network incident prediction, especially ART-style concept-token transformers. The decrypted concept-template field can be used to inspect model inputs and support root-cause analysis, but it should not be passed into the model during training or inference.

Notes

• Samples are grouped by day to support chronological train/validation/test splits.
• The timestamp is stored with UTC+7 timezone information.
• The dataset is already preprocessed with drain3-multiprocess and does not include full raw production logs.