eu-compliance-tools-80k

80K multi-turn tool-calling traces for EU regulatory compliance workflows — GDPR, AI Act, NIS2, DSA, PSD3, CSRD. Each example includes full tool definitions, realistic API calls, error-recovery turns, and role-specific user profiles (DPO, CISO, compliance officer). Apache 2.0 — commercial use permitted.

The only open dataset combining EU regulatory coverage with agentic tool-use patterns. Timed for the EU AI Act enforcement wave.

Quick Load

from datasets import load_dataset

ds = load_dataset("stindardlogic/eu-compliance-tools-80k", split="train")

# Filter by EU regulation
gdpr_ds    = ds.filter(lambda x: x["regulation"] == "gdpr")
ai_act_ds  = ds.filter(lambda x: x["regulation"] == "ai_act")
nis2_ds    = ds.filter(lambda x: x["regulation"] == "nis2")

# Only error-recovery scenarios (model must handle tool failures)
error_ds = ds.filter(lambda x: x["scenario_type"] == "error_recovery")

# DPO role scenarios
dpo_ds = ds.filter(lambda x: "dpo" in str(x.get("_axis_values", "")))

# Access multi-turn conversation
example = ds[0]
print("Regulation:", example["regulation"])
print("Tools used:", example["tools_used"])
for turn in example["conversation"]:
    print(turn["role"], ":", str(turn["content"])[:80])

Schema

{
  "conversation": [
    {"role": "system", "content": "You have access to the following tools: ..."},
    {"role": "user", "content": "Compliance query"},
    {"role": "assistant", "content": null, "tool_calls": [{"name": "query_dpia_registry", "arguments": "..."}]},
    {"role": "tool", "name": "query_dpia_registry", "content": "mock API result"},
    {"role": "assistant", "content": "Compliance guidance response"}
  ],
  "regulation":    "gdpr | ai_act | nis2 | dsa | psd3 | csrd",
  "scenario_type": "single_query | multi_step | error_recovery | cross_regulation",
  "has_call":      true,
  "turns":         9,
  "tools_used":    ["query_dpia_registry", "log_consent_event"],
  "_axis_values":  "Generation metadata (regulation/scenario/user_role)"
}

Coverage

Regulation	Tools covered
GDPR	DPIA registry, consent logging, data subject requests, breach notification
AI Act	Risk classification, conformity assessment, incident reporting
NIS2	Incident reporting, supplier risk scoring, vulnerability disclosure
DSA	Content moderation audit, transparency reports, trusted flagger
PSD3	Payment authorization, SCA checks, fraud monitoring
CSRD	ESG reporting queries, emissions data, supply chain due diligence

Use cases

• Fine-tuning compliance assistants for EU-regulated enterprises
• Benchmarking tool-use accuracy on regulatory workflows
• Building DPO/CISO/compliance officer AI copilots
• EU AI Act conformity testing for agentic AI systems

Generation

Synthetic — generated with local models on Mac Studio M3 Ultra (512 GB). All data (company names, incident IDs, amounts) is fictional. EU regulatory terminology is authentic.

More coming

v0.1. Cross-regulation multi-step chains and non-English EU language versions planned.

💼 Custom EU compliance training data

Need compliance tool-calling data for your specific regulation or jurisdiction? Email contact@stindardlogic.ro