TechTush/swiss-alert-fidelity

Swiss Alert Fidelity — multilingual civil-protection alerts by Apertus 1.5

Evaluation data of the Hack Apertus 2026 Track 2B project Sturmglocke (Team Barry): 30 structured alert cases (natural hazard, zone, invented place names, order, deadline, assembly point, hotline) and 720 alert texts written by Apertus 1.5 8B and 70B in French, Italian, English and Romansh under three conditions, with per-text scores.

Dataset Details

• Curated by: Team Barry (Tushar Agarwal)
• Language(s): French, Italian, English, Romansh (Rumantsch Grischun); German source/facts
• License: CDLA-Permissive-2.0 (instance-level licences in metadata.jsonl)
• Git repository: Sturmglocke (Track 2B submission), track_2b/src/sturmglocke/evals

Uses

Direct Use

Benchmarking whether an LLM preserves safety-critical facts — above all the official hazard term — when producing multilingual public alerts; studying guard pipelines (C2) against plain translation (C0).

Out-of-Scope Use

Not a source of real alerts. Places and municipalities are invented. Not a measure of general translation quality.

Specifications

Conditions: C0 Apertus writes the German alert from the facts, then plain-translates it; C1 slot-first rendering with the official term locked; C2 C1 + deterministic checks + read-back + repair (the shipped pipeline).

Dataset Structure

• eval_cases.jsonl — one structured case per line (case_id, hazard, zone, places, action, deadline, assembly point, hotline …)
• model_responses.jsonl — one alert text per (case, model, condition, language) with: hazard_class (official / acceptable / wrong_process / missing), deterministic checks, romansh_unknown_words (official RG Hunspell), semantic verdicts of two independent open-weights judges (judge_qwen3_235b, judge_llama33_70b), and for Romansh a German backtranslation_de (ZurichNLP romansh-nllb-1.3b) with backtranslation_meaning_ok.
• metadata.jsonl — instance-level source and licence.

Dataset Creation

Curation Rationale

Translated alerts can preserve every number and time but swap the hazard process (debris flow → landslide/rockslide), which changes protective behaviour. We needed a reproducible, model-free way to measure this in Swiss languages.

Source Data

Data Collection

Cases from a seeded generator (seed 42). Responses generated at temperature 0 via the CSCS inference API (Oct 2026).

Source Data Producers and Data Subjects

Structured facts authored by the team; texts produced by Apertus 1.5. No real persons; no data subjects.

Ownership and Consent Management

No personal data; no consent required. Official terminology is referenced by TERMDAT / Pledari Grond entry, single terms only.

Data Processing and Quality Control

Deterministic scoring with an official-terminology glossary (TERMDAT ids, Pledari Grond); Romansh vocabulary with the Lia Rumantscha RG Hunspell dictionary (MIT); semantic read-back by two independent open-weights judges; manual inspection of failure classes (see the project's technical report).

Data Instance Licenses

All instances CDLA-Permissive-2.0 (see metadata.jsonl).

Annotations

Annotation Process

Automatic: deterministic checks + LLM judges (declared). No human labels; judge disagreement concentrates on Romansh, which neither judge reads reliably (documented).

Judge agreement (semantic verdict "meaning preserved")

Languages	n	Agreement	Cohen's kappa
French / Italian / English	540	95.6 %	0.72
Romansh	180	63.9 %	0.22

Annotator Details

Qwen3-235B-A22B-Instruct-2507 and Llama-3.3-70B-Instruct via Hugging Face Inference Providers; ZurichNLP romansh-nllb-1.3b.

Risks, Bias, and Limitations

Invented names; 30 cases; five hazards. Acceptable-synonym lists are limited; unseen correct synonyms count as "missing". Romansh judged by dictionary and back-translation, not by native speakers.

Personally Identifiable Information

None.

Sensitive Information

None.

Recommendations

Use the deterministic checks as the primary metric; treat judge verdicts as secondary, especially for Romansh.