---
pretty_name: Thunderbird
dataset_name: logfit-project/Thunderbird
task_categories:
- text-classification
language:
- en
size_categories:
- 100M<n<200M
annotations_creators:
- logfit-project
license: other
---

# Dataset Card for logfit-project/Thunderbird

## Dataset Summary
The Thunderbird dataset captures console logs from the Thunderbird supercomputer deployed at Sandia National
Laboratories. Entries include alert and non-alert events flagged via category tags, enabling research into
alert detection, failure prediction, and broader log analytics workflows for large-scale clusters.

## Supported Tasks and Leaderboards
- anomaly-detection: distinguish alert-tagged events from routine operational logs.
- log-parsing: template discovery and sequence modeling for large-scale HPC environments.

## Dataset Structure
- `label`: Alert category tag (`-` denotes non-alert informational entries).
- `timestamp`: Unix epoch timestamp for the log event.
- `date`: Calendar date formatted as `YYYY.MM.DD`.
- `user`: Host account name associated with the emitted log.
- `month`: Short month string extracted from the raw message.
- `day`: Day-of-month integer from the message header.
- `time`: Local time-of-day string (`HH:MM:SS`).
- `location`: Combined source identifier conveying host and logger context.
- `component`: Process or subsystem identifier preceding the message body.
- `pid`: Numeric process identifier when provided (set to `-1` if absent).
- `content`: Human-readable message body for the event.
- `anomaly`: Binary indicator (`1` for alert labels, `0` when `label` is `-`).

## Source Data
- **Homepage:** https://github.com/logpai/loghub/tree/master/Thunderbird
- **Original Maintainers:** The LogPAI team (https://logpai.com/).
- **Additional Context:** https://www.usenix.org/cfdr-data#hpc4

## Dataset Creation
Raw Thunderbird logs are streamed line-by-line using a deterministic parser consistent with the published
structured sample. PIDs are extracted from bracketed segments when present, while messages lacking PIDs are
assigned `-1`. Parsing avoids loading the full source file into memory so the complete corpus can be
processed efficiently.

## Uses
Applicable to supervised alert detection, incident trend analysis, failure prediction, and validation of log
parsing algorithms for large distributed systems.

## Citation
- Adam J. Oliner, Jon Stearley. "What Supercomputers Say: A Study of Five System Logs", DSN 2007.
- Jieming Zhu, Shilin He, Pinjia He, Jinyang Liu, Michael R. Lyu. "Loghub: A Large Collection of System Log
  Datasets for AI-driven Log Analytics", ISSRE 2023.

## Dataset Statistics
- Number of log lines: 211212192