Datasets:

---

SlayerLab

/

polish-dynaword

Like 29

Follow

Fabryka AI 101

Description of this PR, first posted on 2026-10-18.

The personal-data pass for uzp_orzeczenia_pl. The shard is unchanged in this PR. src/build_uzp_orzeczenia_pl.py g1 runs once, on a GPU, before the corpus release:

1. NERGAL 2.0.3 (SlayerLab/NERGAL@ee88d626, the revision dynaword-pipeline pins) with its names layer masks person names as [PERSON]. Each ruling's bench and clerk, read from its header, are kept.
2. Rules mask sole-trader and s.c. firm names ([FIRM]) and their street addresses ([ADDRESS]).
3. A sentence that ties a person to a criminal matter or to health becomes [REDACTED:SPECIAL].

No ruling is dropped. The PR also adds:

• the contract test;
• the sources.py entry (chairman, purchaser and provisions are empty in every record);
• datasheet and NOTICE updates;
• a draft legitimate-interest note, data/uzp_orzeczenia_pl/legitimate_interest.md. It is our reading, not a legal opinion.

Measured

320 hand-labelled passages: 240 stratified (seed 112) and 80 more where a firm or special-category cue occurs (seed 1122). One labeller (Claude) did the labels. The person check of 40 passages had not been done when this was posted. Full figures are in artifacts/source_findings.md, the uzp_orzeczenia_pl personal-data entry.

• Person names left at least in part: 3.8% of mentions, weighted by tokens over S1–S4, which partition the shard (bootstrap 0.9–8.9%).
• Firms wholly masked: S1 22/28, S2 21/35, S3 35/57, S4 7/16, S5 162/227, S6 67/83 (seed 1122: S5 118/148, S6 68/139).
• Special-category sentences replaced: S6 15/16 (seed 1122: 11/11). Also replaced although nothing in them is special: 14 sentences (seed 1122: 13).
• Initials are not a target of this pass.
• Over-capture: NERGAL's names model masks party-role nouns ("Odwołujący", "Przystępujący"). In court rulings, 15 of 175 bench mentions are masked.

Release steps (not in this PR)

1. The IPv4 fix (IPV4_RE redacted section numbers; branch uzp-ip-fix, no PR yet) touches the same build script and test. Whichever PR merges second rebases.
2. Run build_uzp_orzeczenia_pl.py g1 --output data/uzp_orzeczenia_pl --nergal <snapshot> on a GPU. It is one process; NERGAL's throughput figures put it at about 9 h on an RTX 4090 (not measured for this pass).
3. Upload with mask_names=false, because this pass has already masked the names.
4. Refresh the stats in artifacts/uzp_orzeczenia_pl_ontology_manifest.json.
5. A NERGAL 2.0.4 that stops masking role nouns is planned. If it is ready before the run, bump NERGAL_VERSION/NERGAL_REVISION and re-measure.

Tests

The full suite was run on this branch and on a clean origin/main worktree (22f9b01):

No test newly fails, and the 21 new tests pass.

🤖 Generated with Claude Code