Datasets:

mischeiwiller
/

gsw-eval

license: cc-by-sa-4.0
language:
  - gsw
  - de
pretty_name: gsw-eval — Swiss-German LID benchmark
size_categories:
  - n<1K
multilinguality: multilingual
annotations_creators:
  - found
source_datasets:
  - mischeiwiller/swiss-german-text
task_categories:
  - text-classification
tags:
  - swiss-german
  - alemannic
  - gsw
  - dialect
  - language-identification
  - low-resource
  - mundart
  - benchmark
configs:
  - config_name: lid
    data_files:
      - split: train
        path: lid_train.jsonl
      - split: test
        path: lid_test.jsonl
  - config_name: probe
    data_files:
      - split: train
        path: probe_gsw_de.jsonl

gsw-eval — Swiss-German evaluation benchmark

A small, honest benchmark over held-out text from mischeiwiller/swiss-german-text. Built by scripts/build_eval_split.py (task 6.1) via the pure mundart.eval layer.

Why not regional dialect-ID? Every row of the source corpus carries region="unknown" — no greenlit source supplies a defensible dialect-region label — so a regional classification task would require inventing labels we cannot defend. The task the data genuinely supports is gsw-vs-de language identification, framed below.

Task 1 — gsw-vs-de language identification (LID)

Given one sentence, predict whether it is written Swiss-German (gsw) or Standard German (de). This is a real, useful gap: off-the-shelf language detectors collapse Swiss-German into German, so a gsw-aware LID number is worth reporting.


Labels	`de`, `gsw` (binary)
Positives (`gsw`)	held-out `swiss-german-text` sentences, sampled round-robin across the three slices (Alemannic-Wikipedia, Tatoeba, Leipzig) so encyclopedic, conversational and web registers are all represented
Negatives (`de`)	the Standard-German side of genuine Tatoeba `gsw→deu` alignment links — real Standard German, never machine-generated
Balance	1:1 — the 334 Standard-German negatives are the binding constraint, matched by an equal gsw sample
Split	per-class stratified, `seed=6`, `test_frac=0.30` → train 468 (234/234), test 200 (100/100); `train ∩ test = ∅`
Primary metric	macro-F1 over `{de, gsw}` (reported with accuracy)

Files: lid_train.jsonl, lid_test.jsonl — one example per line ({id, text, label, source}).

Task 2 — gsw→de understanding probe

The 334 Tatoeba gsw→de aligned pairs (probe_gsw_de.jsonl, {id, gsw, de, source}) ship as a parallel set for a gsw→de understanding probe (retrieval / translation eval). These pairs are kept disjoint from the LID gsw positives — no sentence appears in both — so a model cannot see a probe gsw sentence during LID evaluation.

Scoring

mundart.eval.score(y_true, y_pred) → {n, accuracy, macro_f1, per_class_f1}. The macro-F1 is dependency-free and matches scikit-learn's f1_score(average="macro") over the union of gold ∪ predicted labels.

from mundart.eval import score
# y_true / y_pred are lists of "de" / "gsw"
print(score(y_true, y_pred))

Baseline

A light, CPU-only reference baseline ships with the benchmark: a character n-gram multinomial Naive Bayes (mundart.baseline, orders 2–4, Laplace alpha=1.0). Naive Bayes is chosen over fastText / logreg because it is deterministic by construction — closed-form counts, no optimizer, no RNG, no seed — so one command regenerates the exact number with no thread/seed caveats.

Split	accuracy	macro-F1	F1 `de`	F1 `gsw`
test (n=200)	0.6700	0.6324	0.5147	0.7500

The per-class gap (de < gsw) is honest: the de negatives are short single Tatoeba sentences while the gsw positives include longer Leipzig/Wikipedia text, so the model over-predicts gsw on short inputs. This is a deliberately beatable floor — the benchmark's value is the held-out split + task definition; better LID models should clear 0.63 macro-F1 comfortably.

Regenerate the number (byte-identical — eval/baseline.json carries the full scorecard + config):

# from projects/03-mundart/ (uses only the committed eval/ split)
PYTHONPATH=src .venv/bin/python scripts/run_baseline.py

Reproduce

# from projects/03-mundart/ (needs the gitignored storage/output/*.jsonl slices)
PYTHONPATH=src .venv/bin/python scripts/build_eval_split.py

Regenerating is byte-identical (fixed seed, order-independent stable shuffle). Counts and provenance are mirrored in manifest.json.

Provenance & license

Derived from mischeiwiller/swiss-german-text (per-slice licenses: Alemannic-Wikipedia CC-BY-SA-4.0, Tatoeba CC-BY-2.0-FR, Leipzig CC-BY-4.0).
The benchmark inherits the most restrictive compatible license of its included text (CC-BY-SA-4.0); eval code is Apache-2.0.
PII was scrubbed at ingest and re-scrubbed (idempotent) when the split is built.

Companion artifacts

Part of the Mundart project (dataset + eval + demo):

Dataset: mischeiwiller/swiss-german-text — the open written-gsw corpus this benchmark is carved from.
Space: mischeiwiller/mundart-explorer — paste a sentence → gsw-vs-de LID guess (this baseline) + nearest gsw→de aligned pair from the probe set.
Neighbour reference: ZurichNLP/swissbert — a Swiss-German-aware multilingual encoder; a strong candidate to evaluate on this LID task.

Status

Task 6.1 shipped the held-out split + task/metric definitions; task 6.2a adds the deterministic char n-gram NB baseline (macro-F1 0.6324) + the reproducible command above. Task 6.2b pushes mischeiwiller/gsw-eval (private) with this card, the split, the probe, and baseline.json.