llm-training-alert-trace Dataset

📖 Overview

llm-training-alert-trace is an open-source dataset comprising anonymized production-scale alert traces from large-scale Large Language Model (LLM) training jobs. It captures the interplay between system alert events and node health status in high-performance computing environments.

This dataset is designed to foster research in:

• AIOps & Reliability Engineering: Automated fault diagnosis and root cause analysis.
• System Characterization: Understanding failure patterns in distributed training.

📄 Paper

This dataset is associated with the following paper:

Reflex: Event-Driven Automated Fault Localization for Large-Scale LLM Training
Hua Ding, Yun Zhang, Bo Zhang, Wenxiao Wang, Libo Chen, Huan Yu, Zhe Nan, Zuquan Song, Weiqiang Lou, Gaohong Liu, Xi Yang, Yuhan Li, Qinlong Wang, Shuguang Wang, Wencong Xiao, and Shenghong Li.
In 34th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (FSE Companion '26), July 5-9, 2026, Montreal, QC, Canada. ACM, New York, NY, USA, 13 pages.
https://doi.org/10.1145/3803437.3805245

🤗 Quick Start with Hugging Face

from datasets import load_dataset

# Load the job metadata (default configuration)
jobs = load_dataset("bytedance-research/llm-training-alert-trace", "jobs", split="train")

# Load the alert events
events = load_dataset("bytedance-research/llm-training-alert-trace", "events", split="train")

The nested JSONL representation can be downloaded directly from data/llm_alert_trace.jsonl.

📂 Dataset Structure

The dataset is available in multiple formats for convenience:

Original Format (Example only)

root/
├── Job1/
│   ├── meta.json           # Job metadata with IP status
│   └── event_report.csv    # Time-series alert events

Note: Only Job1 is provided as an example of the original data format. The full dataset is available in JSONL and Parquet formats.

JSONL Format (Recommended)

Single file: data/llm_alert_trace.jsonl

Parquet Format

Two files:

• data/jobs.parquet - Job metadata
• data/events.parquet - Event data

📊 Data Source & Conversion

Original Data Source

The original dataset contains 98 training job directories (Job1-Job98), each containing:

• meta.json: Job metadata including start/end times, GPU generation, and node health status
• event_report.csv: Time-series alert events with anonymized IP addresses

Data Conversion Process

We provide two conversion scripts to transform the original data into more usable formats:

1. JSONL Conversion: Combines metadata and events into a single JSONL file for easy exploration
2. Parquet Conversion: Splits data into two Parquet files for efficient columnar storage

The conversion scripts preserve all original data while providing more accessible formats for research and analysis.

📝 Data Formats

JSONL Format

Each line represents a complete training job:

{
  "job_id": "Job1",
  "start_time": "1969-10-15 13:49:13",
  "end_time": "1969-10-15 14:11:55",
  "gpu_generation": "Hopper",
  "ip_status": {
    "47d7:25d3:c729:ae5a:d0c8:a53:1a9:ca8f": "normal"
  },
  "events": [
    {
      "ip": "ee23:95e7:f4cb:bed:e641:52f6:5c1:4823",
      "event_time": "1969-10-15 13:49:24",
      "event": "failed_to_forward"
    }
  ]
}

Parquet Format

jobs.parquet:

Field	Type	Description
job_id	String	Unique job identifier
start_time	String	Anonymized start timestamp
end_time	String	Anonymized end timestamp
gpu_generation	String	GPU generation
ip_status	String	JSON-encoded node health labels keyed by pseudonymous IPv6-form tokens

events.parquet:

Field	Type	Description
job_id	String	Job identifier (foreign key)
ip	String	Anonymized IPv6 address
event_time	String	Alert timestamp
event	String	Alert event type

🛠️ Conversion Scripts

If you need to regenerate the converted formats:

# Convert to JSONL
python scripts/convert_to_jsonl.py

# Convert to Parquet
python scripts/convert_to_parquet.py

🛡️ Data Anonymization & Privacy

To reduce exposure of sensitive production information, direct identifiers have been removed or transformed:

• IP Addresses: Replaced with pseudonymous IPv6-form tokens. Tokens may recur across jobs and should not be treated as unlinkable.
• Timestamps: For each job, all timestamps (including start_time, end_time, and event_time) are randomly shifted forward to obscure absolute production windows, while preserving the relative time intervals within each job.
• Sensitive Identifiers: User IDs, specific cluster names, and proprietary error messages have been removed or tokenized.

📄 License

CC BY 4.0 - Attribution 4.0 International License

📚 How to Cite

When using this dataset, please cite the following paper:

@inproceedings{ding2026reflex,
  author = {Ding, Hua and Zhang, Yun and Zhang, Bo and Wang, Wenxiao and Chen, Libo and Yu, Huan and Nan, Zhe and Song, Zuquan and Lou, Weiqiang and Liu, Gaohong and Yang, Xi and Li, Yuhan and Wang, Qinlong and Wang, Shuguang and Xiao, Wencong and Li, Shenghong},
  title = {Reflex: Event-Driven Automated Fault Localization for Large-Scale LLM Training},
  year = {2026},
  isbn = {979-8-4007-1145-5/26/07},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3803437.3805245},
  doi = {10.1145/3803437.3805245},
  booktitle = {34th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (FSE Companion '26)},
  location = {Montreal, QC, Canada},
  series = {FSE Companion '26}
}