Devseis/devseis-ai-act-classifier-v5-data

Devseis AI Act Classifier — v5 training and evaluation data

Newer version: Devseis/devseis-ai-act-classifier-v6-data extends this data with 39 scenarios and a second writing style for every training example.

A Devseis dataset (data_v5_systems). Used to train Devseis/devseis-ai-act-classifier-v5, which powers Caveat.

This is the data the v5 classifier was trained and evaluated on. It is published so the numbers in the model card and on the Caveat page can be checked. It is synthetic, and it is not legal advice.

Files

File	What it is	Rows	Scenarios
data/train_v5.csv	training set (text, label)	636	212
data/val_v5.csv	validation set, used only to choose settings	147	49
data/test_v5.csv	test set, used once per final configuration	198	66
data/scenarios_v5.csv	one row per distinct scenario (text, label, legal_basis)	327	327
CHANGES_v3.md, CHANGES_v4.md, CHANGES_v5.md	what changed in each version and why

Labels: prohibited, high_risk, limited_risk, minimal_risk. Counts by scenario:

Split	high_risk	minimal_risk	prohibited	limited_risk
train	90	60	38	24
val	19	14	9	7
test	27	19	11	9

Loading

from datasets import load_dataset
ds = load_dataset("Devseis/devseis-ai-act-classifier-v5-data")              # train / validation / test
scenarios = load_dataset("Devseis/devseis-ai-act-classifier-v5-data", "scenarios")

How it was made

• Synthetic, LLM-written scenarios. Each is based on the text of Regulation (EU) 2024/1689 (Art 5, Art 50, Annex I, Annex III), its recitals, and the European Commission's draft guidelines on classifying high-risk AI systems.
• Template-expanded. Each scenario appears as 3 rows. The rows differ only in an opening phrase ("Our client wants to deploy…", "A startup is building…"). In about half the scenarios, one of the three rows has no opening phrase. The real sample size is the scenario count, not the row count.
• Split by scenario. All rows of a scenario sit in the same split. No scenario appears in more than one split. test_v5 is test_v4 plus 5 new scenarios, so results can be compared with the v4 and earlier numbers on the same rows.

Known caveats

• Every scenario is synthetic, and the test sets are small (66 scenarios, 11 of them prohibited). Scores on this data are an upper bound on real-world performance.
• Age and sex estimation from biometrics is labelled high_risk (Annex III(1)(b)) as a consistent policy choice, because the law is unsettled on it.
• The AI Omnibus scenarios (non-consensual intimate imagery and AI-generated child sexual abuse material) are labelled prohibited, which is the law from 2 December 2026. They are one-line, non-graphic descriptions of banned systems. Their legal_basis citations, Art. 5(1)(ba) and (bb), are verified against the Official Journal text of the AI Omnibus, Regulation (EU) 2026/1744, Art. 1(7). Under the new Art. 5(1a), a system is banned only where that output is its intended purpose, or a reasonably foreseeable outcome without adequate safeguards. All 9 scenarios describe systems built for that purpose.
• The CHANGES_*.md files were written for the build folders, so their relative paths (such as ../relabel_audit/) point to files that are not published here.

License

CC-BY-4.0.