Snaseem2026/iac-drift-pairs

IaC Drift Pairs

Declared Infrastructure-as-Code state → drifted live state → structured diff, cause, risk, and fix.

Configuration drift — the gap between what your Terraform/Pulumi/CloudFormation declares and what actually exists in the cloud — is one of the most expensive and security-critical problems in cloud operations. Yet there is no public, labeled dataset that pairs a declared state with its drifted counterpart and the root cause, risk, and reconciliation. This dataset fills that gap.

Every row is generated by chaos-style drift injection: a resource is declared, a specific drift is applied, and the drifted state is observed. Because the drift is injected deliberately, every label is ground truth — no manual annotation, no scraping, no licensing ambiguity.

Why this is useful

• Trainable task: given declared_state (+ declared_hcl) and observed_state, predict the drift_diff, the risk, and the reconciliation. Directly usable to train IaC copilots, drift detectors, and policy-as-code assistants.
• Security-relevant: flagship scenarios cover world-open security groups, disabled S3 encryption, public ACLs, and IAM privilege escalation.
• Fully reproducible: deterministic generator (seeded). Re-run it and get the exact same rows.

Schema

Field	Type	Description
id	string	Stable row id (drift-000123)
template_id	string	Drift scenario template
resource_type	string	e.g. aws_security_group, aws_s3_bucket, aws_db_instance, aws_lambda_function_url, kubernetes_pod, aws_iam_policy (16 scenarios)
resource_name	string	Generated resource name
declared_state	json	Intended configuration (IaC)
observed_state	json	Actual live configuration (drifted)
declared_hcl	string	declared_state rendered as Terraform HCL
drift_diff	json	Field-level diff (added / removed / modified)
num_changes	int	Number of drifted fields
drift_cause	string	e.g. manual_change_via_console, emergency_hotfix, autoscaling_policy_override
risk	json	{category, severity} — security / cost / availability / compliance
reconciliation	json	The action + HCL patch that closes the drift
narrative	string	Human-readable description of the incident
postmortem	string	Full structured post-mortem (only when generated with --enrich)
observed_source	string	simulated or live_localstack

Example row (abridged)

{
  "id": "drift-000042",
  "resource_type": "aws_security_group",
  "declared_state": {"ingress": [{"from_port": 22, "cidr_blocks": ["10.0.0.0/16"]}]},
  "observed_state":  {"ingress": [{"from_port": 22, "cidr_blocks": ["10.0.0.0/16"]},
                                  {"from_port": 22, "cidr_blocks": ["0.0.0.0/0"]}]},
  "drift_diff": [{"path": "ingress[1]", "change": "added",
                  "observed": {"from_port": 22, "cidr_blocks": ["0.0.0.0/0"]}}],
  "risk": {"category": "security", "severity": "critical"},
  "reconciliation": {"action": "remove_ingress_rule", "target_path": "ingress[1]"}
}

How it was built

Two backends produce the identical schema:

• simulate (default): pure-Python, no dependencies. Instant, high-volume.
• live: applies each declared resource to LocalStack, injects the drift through the real AWS API, and reads the drifted state back — so observed_state reflects provider-accurate behaviour.

pip install -r requirements.txt

# fast, no Docker
make generate NUM=2000

# real drift from LocalStack
make up
make generate-live NUM=200

Load it

import json
from datasets import load_dataset

ds = load_dataset("Snaseem2026/iac-drift-pairs")
row = ds["train"][0]

# nested fields are stored as JSON strings for a stable, viewer-friendly schema
declared = json.loads(row["declared_state"])
observed = json.loads(row["observed_state"])
diff     = json.loads(row["drift_diff"])
risk     = json.loads(row["risk"])

A runnable walkthrough (load → inspect a drift → plot the risk/cause distributions) is in examples/quickstart.ipynb.

Suggested benchmark tasks

1. Drift detection — classify whether observed_state drifted from declared_state.
2. Root-cause / risk classification — predict drift_cause and risk.
3. Reconciliation generation — generate the HCL patch that fixes the drift.

A runnable baseline (TF-IDF + logistic regression) ships in benchmark/:

pip install scikit-learn
python benchmark/run_benchmark.py --data data

Reference numbers on 3,000 simulated rows (seed 7):

Task	Accuracy	Macro-F1
drift_detection	1.00	1.00
risk_severity	1.00	1.00
risk_category	1.00	1.00
drift_cause	0.47	0.48

drift_cause is deliberately hard — the same drift can have several plausible causes — leaving real headroom for fine-tuned models. Reconciliation generation is a generative task not covered by this baseline.

Optional: full post-mortems

Each row can carry a structured incident post-mortem (Summary / Detection / Root cause / Impact / Remediation / Prevention):

python -m generator.build_dataset --num 3000 --enrich template   # deterministic
python -m generator.build_dataset --num 3000 --enrich llm         # needs OPENAI_API_KEY

Publish to the Hub

pip install huggingface_hub
huggingface-cli login
python scripts/push_to_hub.py --repo Snaseem2026/iac-drift-pairs

Limitations

• Resource coverage spans 16 high-value scenarios across AWS (SG, S3, EC2, RDS, Lambda, VPC, CloudWatch, IAM, ECR) and Kubernetes. Extend generator/scenarios.py to add more.
• simulate rows use hand-modeled provider behaviour; live rows use LocalStack, which approximates but is not identical to real AWS.

License

Apache-2.0. All data is synthetically generated — no third-party content.

Citation

@misc{naseem2026iacdriftpairs,
  title        = {IaC Drift Pairs},
  author       = {Naseem, Sharyar},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Snaseem2026/iac-drift-pairs}}
}

A machine-readable CITATION.cff is included in the repository.