eu-compliance-tools-80k

78,547 multi-turn English conversations of a compliance officer or developer interacting with an LLM assistant that has access to EU-regulation APIs. Covers GDPR, EU AI Act, PSD3, DSA, CSRD, NIS2, including realistic API errors, retries, and permanent-failure paths.

Why this exists

The EU AI Act enforcement spike on 2 August 2026 plus the broader EU regulatory stack (GDPR, PSD3 / Open Banking, DSA, CSRD, NIS2) are driving enterprise demand for LLMs that don't just discuss compliance APIs but actually use them — query DPIA registries, log consent events, classify AI systems by risk, generate Annex IV dossiers, validate Strong Customer Authentication, file incident reports.

Existing tool-calling datasets (Glaive v2, xLAM-60k) cover generic APIs and happy-path flows. This dataset covers the EU-regulatory tool surface specifically, with explicit error-recovery and partial-success scenarios that real compliance workflows hit constantly.

Schema

{
  "conversation": [
    {"role": "system", "content": "You have access to: query_dpia_registry, log_consent_event, ..."},
    {"role": "user", "content": "Is processing activity DPIA-123 registered? If yes, log a consent event for user X."},
    {"role": "assistant", "content": "<functioncall>{\"name\":\"query_dpia_registry\",\"arguments\":{\"id\":\"DPIA-123\"}}</functioncall>"},
    {"role": "tool", "content": "{\"registered\": true, \"data_categories\": [...], \"last_review\": \"2026-04-15\"}"},
    {"role": "assistant", "content": "<functioncall>{\"name\":\"log_consent_event\",\"arguments\":{\"user_id\":\"X\",...}}</functioncall>"},
    {"role": "tool", "content": "{\"status\":\"ok\",\"event_id\":\"...\"}"},
    {"role": "assistant", "content": "Done. The activity is registered (last reviewed 2026-04-15) and I've logged the consent event per GDPR Article 7."}
  ],
  "regulation": "gdpr",
  "scenario_type": "single_query",
  "has_call": true,
  "turns": 9,
  "tools_used": ["query_dpia_registry", "log_consent_event"]
}

Coverage

Axis	Values
regulation	gdpr, ai_act, psd3, dsa, csrd, nis2
scenario_type	single_query, chained_lookup, data_subject_request, incident_response, audit_preparation, compliance_check, reporting_obligation, cross_border_transfer
failure_mode	no_failure, retry_succeeds, permanent_failure, partial_success
user_role	dpo, developer, compliance_officer, product_manager, external_auditor, legal_counsel

Approximately 25% of conversations include failure_mode != no_failure — realistic API errors with retry logic or graceful permanent-failure handling. This is the key differentiator from happy-path-only tool-calling corpora.

Function names follow each regulation's domain naturally:

• GDPR: query_dpia_registry, log_consent_event, anonymize_dataset, check_data_residency, validate_dsar_request
• AI Act: classify_ai_system_risk, register_high_risk_system, check_prohibited_use, generate_annex_iv_dossier
• PSD3: validate_strong_customer_auth, check_open_banking_consent, log_payment_initiation
• (similar for DSA, CSRD, NIS2)

Article numbers cited in assistant summaries are real (e.g., "per GDPR Article 30", "per AI Act Annex IV section 3").

Training recipes

SFT

Use the conversation array directly with your usual chat-template format. The dataset is already structured as OpenAI-compatible tool-call messages.

Specialized fine-tunes

Filter by regulation for vertical models (e.g., a GDPR-only Mistral-7B), or by scenario_type for task-specific tuning (e.g., incident_response specialist).

Benchmark

Use failure_mode == permanent_failure records as a robustness eval: does your model fail gracefully and escalate appropriately?

Independent quality evaluation

This dataset was independently evaluated by a separate judge model (different from the Critic used during generation) to give an honest external quality signal.

Judge model: DeepSeek-R1 70B (local)
Sample: 198 records stratified by regulation
Method: structured per-record verdict (pass/fail) on 4-7 schema-specific checks

Headline numbers

• Pass rate: 47.0% (93/198)
• Fail rate: 52.5% (104/198)
• Unparseable judge output: 1 (0.5%)

Per regulation breakdown

regulation	Pass	Fail
gdpr	19/33 (58%)	14/33 (42%)
ai_act	19/33 (58%)	14/33 (42%)
nis2	17/33 (52%)	15/33 (45%)
csrd	16/33 (48%)	17/33 (52%)
dsa	15/33 (45%)	18/33 (55%)
psd3	7/33 (21%)	26/33 (79%)

Most common failure modes

• article_numbers_real — 67/198 (34%) records fail this check
• call_matches_schema — 57/198 (29%) records fail this check
• system_schema_valid — 31/198 (16%) records fail this check
• no_call_consistent — 13/198 (7%) records fail this check
• final_summary_coherent — 12/198 (6%) records fail this check

Guidance

🔴 Substantial quality issues. Treat as research-only. Filter aggressively before production use. See 'Limitations' section below for known issues.

Reproducibility

The judge prompts and scripts are open: ask via contact@bidbrief.ro for the full evaluation script + per-record verdicts JSONL. We can also re-run with a different judge model on request.

How it was built

Pipeline: Generator → Teacher → Critic → Formatter.

• Generator combinatorially explores the diversity axes
• Teacher generates the conversation (multi-step, including tool calls + tool responses + final summary)
• Critic validates: function-name domain match, article-number authenticity, error-response realism, no fabricated case law, no personal data

All inference on Apple Mac Studio M3 Ultra (512 GB unified memory) — open-weight models, no cloud API dependency. Generation time: ~3.5 hours for the full 80K. Acceptance rate: 98.2% (78,547 accepted, 1,453 rejected by the critic on regulation-specific validation rules — stricter than typical synthetic-data pipelines).

Use cases

• Fine-tune EU-fintech / regtech / govt-AI LLMs to use compliance APIs natively
• Train internal-tools assistants for DPOs and compliance teams
• Benchmark agentic LLMs on error-recovery (often the missing capability)
• Distil a smaller specialized model from a larger reasoning model on this corpus

License

CC-BY-4.0 — free for any use including commercial fine-tuning. Attribution: cite this dataset card.

Commercial use & custom domain extensions

For domain-specific extensions — e.g., audited compliance tier with EUR-Lex citation verification, your specific Open Banking API surface, your CMS Article 30 register schema, NIS2 Annex IV reporting templates tailored to your sector — contact contact@bidbrief.ro.

Commercial license tier covers: audited datasets with human-verified article citations, custom regulation schema additions, and bundled fine-tuned model SKUs (Mistral / Qwen base + LoRA on this corpus).

Honest disclosure

• Synthetic dataset — does not replace legal advice. Tool call patterns reflect plausible API surfaces but specific compliance interpretations should be verified against current EUR-Lex text + your DPO / legal counsel.
• No real personal data (CNP, IBAN, email, real names) — all examples use placeholders.
• No fabricated CJEU / national-court case law — assistant summaries reference real article numbers but avoid specific case citations.
• 1.8% of generated examples were rejected by the critic for failing one of the validation rules — see eu-compliance-tools-80k.rejections.jsonl (gitignored from upload, available on request for transparency on rejection patterns).

Author

Built by Stindard Logic — synthetic dataset and fine-tuned model SKUs for EU/RO regulatory AI. Open to collaboration with the OpenLLM-Ro consortium, European sovereign-LLM initiatives, and EU compliance vendors.

Citation

@misc{eu-compliance-tools-80k-2026,
  author = {Stindard Logic},
  title  = {eu-compliance-tools-80k: Multi-turn EU compliance tool-calling dataset (GDPR + AI Act + PSD3 + DSA + CSRD + NIS2)},
  year   = {2026},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/datasets/stindardlogic/eu-compliance-tools-80k}
}