pl-temporal / README.md
bartoszkobylinski1's picture
Upload README.md with huggingface_hub
bae83e5 verified
|
Raw
History Blame Contribute Delete
4.44 kB
metadata
license: cc-by-4.0
language:
  - pl
task_categories:
  - question-answering
tags:
  - legal
  - temporal-reasoning
  - polish-law
  - benchmark
pretty_name: PL-Temporal Track A
size_categories:
  - n<1K

PL-Temporal — Track A (Diagnostic Challenge Set v0.2)

A diagnostic challenge set for closed-book temporal QA over Polish statutes: which version of the law was in force on a given date, when did an act enter into force, which Dziennik Ustaw position is the current consolidated text. 80 questions across 64 acts (79 scored, 1 quarantined after human legal review), 4 question families, deterministic string-nugget scoring — no LLM-as-judge.

Canonical repository (scoring scripts, results, review protocol): https://github.com/bartoszkobylinski/pl-temporal

Files

file content
questions.json 80 questions + gold answers + ELI act identifiers + a _CANARY_ record
nuggets.json scoring nuggets (exact / regex), one implementation of matching in the repo
quarantine.json items excluded from scoring, with evidence

Question families

family n what it tests
A1 24 commencement date of an act, derived from its commencement clause
A2 8 which Dz.U. position is the consolidated text in force on a date
A3 24 whether an act was in force on a given date (TAK/NIE)
A4 24 whether an act was amended in a given period (TAK/NIE)

v0.2 results (2026-08-14, 79 scored items; mean over valid draws)

Three metrics — no single leader without a utility function. Coverage: how often the model answers. Selective accuracy: how often it is right when it answers. End-to-end: usefulness when every question needs an answer.

model N valid coverage selective acc end-to-end abstained/draw
gpt-5.6-sol 5 100% 69.9% 69.9% 0
claude-opus-5 5 85.1% 74.2% 63.0% 11.6
gpt-5.6-terra 5 100% 57.2% 57.2% 0
gemini-3.1-pro 3 82.3% 64.6% 53.2% 14.0
gemini-3.5-flash 2 93.7% 56.8% 53.2% 5.0
pllum-12b 5 100% 44.3% 44.3% 0
bielik-11b-v3 5 100% 43.0% 43.0% 0

Majority baselines for the binary families: A3 = 65.2%, A4 = 50.0%. v0.2 is frozen as a diagnostic challenge set (no-signal pruning used v0.1 model results); a frozen v1.0 with a pre-registered selection procedure is planned. Full findings, limitations and incident log: RESULTS-v0.2.md in the canonical repo.

Gold answers for every analyzer-flagged item were verified by the dataset owner against ISAP commencement clauses (2026-08-14): zero gold edits, one ambiguous item quarantined; an independent second annotation pass is planned for v1.0. Notable finding: zero-abstention is a property of vendor training decisions, not of model size — both GPT-5.6 models abstained 0/400, matching the Polish models, while legacy gpt-5 was the heaviest abstainer in the roster.

Contamination canary

questions.json contains a _CANARY_ record with a unique GUID that is not a question and must never be sent to models during evaluation. If a model can reproduce this GUID, this dataset was in its training data — please report such a finding. Note the asymmetry: reproducing the GUID is positive evidence of contamination, but failing to reproduce it does not prove the absence of contamination.

Data provenance and legal basis

Source of legal texts and metadata: Internetowy System Aktów Prawnych (ISAP), Kancelaria Sejmu RP — https://isap.sejm.gov.pl — and the Sejm ELI API — https://api.sejm.gov.pl / https://eli.gov.pl. Retrieved: 2026-08-12/13. The texts have been processed (fragment extraction, paraphrase, derived dates); they are informational only and are not a source of law. Authentic texts are solely those promulgated in Dziennik Ustaw / Monitor Polski. This dataset is not legal advice.

Statute texts are excluded from copyright under art. 4 of the Polish Copyright Act and are in the public domain; the CC BY 4.0 license applies to the compilation, questions, nuggets, and metadata. Attribution required — cite this dataset and state the ISAP/ELI source. Provided as-is, no warranties (CC BY 4.0 Section 5).

Citation

Kobylinski, B. (2026). PL-Temporal: a benchmark of temporal legal reasoning over
Polish statutes (v0.2). https://github.com/bartoszkobylinski/pl-temporal