CoolFace
Modelpublic

tseidl/eu-acquis-regulatory-tone

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face
0likes6downloads
Model Card

eu-acquis-regulatory-tone

Regulatory tone of an EU legislative article: obligation, prohibition, permission, or none.

Fine-tuned ModernBERT-large encoder from the paper Mapping Europe's Digital Acquis (Seidl & Kosti, 2026). The model classifies individual provisions of EU regulations, directives, and decisions — recitals and (sub-)articles as extracted with eurlex-builder — and was trained on labels produced by a five-model LLM ensemble whose coding instructions were validated against two human coders.

Labels

obligation, prohibition, permission, none. Applies to: articles only (recitals carry no deontic force).

Training

  • —Base model: answerdotai/ModernBERT-large (ModernBertForSequenceClassification); max length at training 1024 tokens.
  • —Training data: provisions labelled by the LLM ensemble (majority vote of Qwen3-235B, Llama-3.3-70B, gpt-oss-120b, GPT-5.2, DeepSeek-V4 Pro); soft labels (per-class vote fractions); negative:positive ratio 4:1.
  • —Hyperparameters: learning rate 2e-05, weight decay 0.1, 5 epochs.
  • —Saved 2026-06-02.

Evaluation

  • —Held-out distillation test (agreement with the ensemble's labels): F1 = 0.936.
  • —Against the human-coded validation set (n = 292, training items excluded by id): agreement 0.918, Krippendorff's α = 0.862.
  • —Same, additionally excluding validation items whose exact text recurs in the training pool (n = 272): α = 0.855.
  • —For reference: the LLM ensemble reaches α = 0.883 against the same coder; the two human coders agree at α = 0.903.

Full validation design, error rates at natural prevalence, temporal stability, prompt stability, and interpretability checks are in the paper's online appendix.

Intended use and limits

  • —Input: one provision of EU secondary legislation in English (the text field as produced by eurlex-builder). The model was trained on provisions of 50 or more characters; very long provisions were truncated at 1,024 tokens during training.
  • —Scope: articles only (recitals carry no deontic force). Applying it to other document types (communications, national law, case law) is untested.
  • —Classification error is documented but not zero; for aggregate analyses, see the paper's discussion of error rates and their correction.

Usage

python
from transformers import pipeline
clf = pipeline("text-classification", model="tseidl/eu-acquis-regulatory-tone", truncation=True, max_length=1024)
clf("Member States shall ensure that all consumers have access to an adequate broadband internet access service at an affordable price.")

Citation

Seidl, T. & Kosti, N. (2026). Mapping Europe's Digital Acquis: A Granular History of EU Digital Policymaking. Preprint.

Corpus construction: eurlex-builder (https://github.com/tseidl/eurlex-builder; archived at https://doi.org/10.5281/zenodo.21496963).

Model card generated 2026-09-16 from the training metadata and the paper's validation results file.