ITBench-Lite Dataset Card

Dataset Overview

Dataset Name: ITBench-Lite
Organization: IBM Research
License: Apache 2.0
Language: English
Paper: ITBench: Evaluating AI Agents across Diverse Real-World IT Automation Tasks
GitHub: ITBench

ITBench-Lite is a systematic framework for benchmarking LLMs and AI Agents on real-world IT automation tasks. This dataset contains 50 scenarios across two critical domains:

• Site Reliability Engineering (SRE): 35 scenarios with environment snapshots for incident diagnosis
• Financial Operations (FinOps): 15 scenarios with synthetic cost anomaly data for spend analysis

Summary

ITBench evaluates LLMs against reasoning and decision-making for IT automation tasks across three domains:

1. Site Reliability Engineering (SRE): Ensures app resilience and performance through incident diagnosis
2. Financial Operations (FinOps): Manages IT spend by performing cost anomaly analysis
3. Compliance & Security Operations (CISO): Manages threats and assesses policies by conducting compliance posture assessment (coming soon)

This leaderboard represents a subset of the broader ITBench and includes selected SRE and FinOps scenarios in offline environments. We will continue enriching the scenario set and incorporating agentic evaluations into the leaderboard over time.

Key Resources

• Paper
• GitHub
• Kaggle Leaderboard

Dataset Structure

Directory Organization

ITBench-Lite/
├── snapshots/
│   ├── sre/
│   │   └── v0.2-B96DF826-4BB2-4B62-97AB-6D84254C53D7/
│   │       ├── Scenario-1/
│   │       │   ├── alerts/                     # Alert snapshots over time
│   │       │   ├── metrics/                    # Prometheus metrics
│   │       │   ├── ground_truth.yaml          # Ground truth entities
│   │       │   ├── k8s_events_raw.tsv         # Kubernetes events
│   │       │   ├── k8s_objects_raw.tsv        # Kubernetes resources
│   │       │   ├── otel_logs_raw.tsv          # OpenTelemetry logs
│   │       │   └── otel_traces_raw.tsv        # OpenTelemetry traces
│   │       ├── Scenario-2/
│   │       └── ... (35 scenarios total)
│   │
│   └── finops/
│       └── v0.1-finops-anomaly/
│           ├── Scenario-1/
│           │   └── anomaly.json               # Cost anomaly details
│           ├── Scenario-2/
│           └── ... (15 scenarios total)

Data Files by Domain

SRE Scenarios (35 scenarios)

Each SRE scenario contains observability data from orchestrated environments where faults were injected:

File	Format	Description
alerts/	JSON	Time-series alert snapshots from monitoring systems
metrics/	Prometheus	Infrastructure and application metrics
ground_truth.yaml	YAML	Ground truth faulty entities and metadata
k8s_events_raw.tsv	TSV	Kubernetes cluster events
k8s_objects_raw.tsv	TSV	Kubernetes resource states (pods, services, deployments)
otel_logs_raw.tsv	TSV	Application and system logs via OpenTelemetry
otel_traces_raw.tsv	TSV	Distributed traces via OpenTelemetry

FinOps Scenarios (15 scenarios)

Each FinOps scenario contains synthetic cost anomaly data:

File	Format	Description
anomaly.json	JSON	Cost anomaly details (date, account_id)

Example anomaly.json:

{
    "date": "2025-10-13",
    "account_id": "49tAGZz"
}

Tasks

SRE: Fault Localization

Identify the faulty entity or resource (e.g., a specific pod, service, or deployment) that caused the incident based on logs, traces, metrics, Kubernetes events and resources from the environment snapshot.

FinOps: Cost Anomaly Analysis

Identify the resource responsible for anomalous cost changes using synthetic cost anomaly details and workload configuration.

CISO: Compliance Assessment (Coming Soon)

Generate the correct Kyverno policy (Kubernetes) or OPA policy (RHEL) based on regulatory rules and natural-language Kubernetes/RHEL configurations.

Limitations

The SRE scenarios are exported snapshots from sandboxed live Kubernetes environments. While sourced from real running systems, the snapshot format inherently lacks the dynamic characteristics of live environments:

• Real-time observability (streaming metrics, logs, traces)
• Runtime state changes and non-deterministic behavior
• Interactive debugging and human-in-the-loop investigation
• Active remediation capabilities (deployments, rollbacks, scaling)

The static nature of these exports makes them suitable for diagnostic analysis but does not capture the full complexity of live incident response.

Related Datasets

• ITBench-Trajectories: Complete agent execution traces with reasoning steps, tool usage, and evaluation metrics for 105 trajectories across 35 SRE scenarios

Citation

If you use ITBench in academic or industrial work, please cite:

@misc{jha2025itbench,
  title={ITBench: Evaluating AI Agents across Diverse Real-World IT Automation Tasks},
  author={Jha, Saurabh and Arora, Rohan and Watanabe, Yuji and others},
  year={2025},
  eprint={2502.05352},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2502.05352}
}

Contact & Support

• GitHub Issues: ITBench Issues
• Kaggle Discussion: ITBench Competition
• Paper: arXiv:2502.05352

Last Updated: January 2026