Datasets:
Add a pinned NERGAL 2.0.1 post-pass for built shards (v2#60)
#96
by ppuzio - opened
Prep for the NERGAL pass (slayerlabs/polish-dynaword-v2#60). No shard is changed; the pass itself runs after the v0.2.6 freeze, one source per PR.
src/nergal_pass.py <key>runs the built shard throughSlayerLab/NERGAL2.0.1, pinned to revision960c515d96d34087ef1159c868a9d72e138834ab(tag v2.0.1). Every file in the snapshot (Hub download or--snapshot DIR) is checked against a sha256 pin (the same pins as dynaword-pipeline'snergal_snapshot.py).- It rewrites
text, recountstoken_count(changed rows only), and updatesstats.json(tokens/characters) andsample.jsonl, keeping each file's format and line ending. It writes<key>.nergal.json: version, revision, options, device/dtype, counts, Parquet sha256 before/after. A shard that already has a record is refused. The sidecar formats are checked before the scrub, so a long run can't fail at the end. - Not rewritten: attribution
text_sha256, manifests/checksums, card excerpts. The source's contract tests catch the stale ones; they get fixed in that source's PR along with the card's Privacy section andlegal_note. - Placeholders: NERGAL ≥ 2.0.0 writes
[PHONE], but AGENTS.md said[Telefon](the pre-NERGAL scrubbers' tag; on main it appears in ted_pl 785×, sejm_interpellations 320×, kultura 12×). The pass renames[Telefon]→[PHONE]so a released shard uses one phone tag, and AGENTS.md now says[PII]/[PHONE]. src/nergal_requirements.txt: the pinned torch/transformers/tiktoken/pyarrow.
Checks:
src/test_nergal_pass.py: 10 tests with a stand-in scrubber (no torch).- Full suite: 713 passed.
- Smoke run with the real pinned model (MPS, float32) on a synthetic 3-row shard: e-mail →
[PII], phones →[PHONE],[Telefon]→[PHONE], the MP's name left in place. One miss: intel. 601 222 333, PESEL 44051401359the PESEL came out[PHONE]. The NERGAL card puts PESEL under[PII]and the rules give[PII]; the model labels the number a phone in a contact line (0.998) and the union ranks phone above pii. Alone,PESEL 44051401359.gives[PII]. A NERGAL bug, not policy. render_sample.py: not applicable, no shard changed.
Known limits: one process, and the whole Parquet in memory (fine up to plwiki_user_talk's 380k rows). Masking can create new exact duplicates inside a source; dedup is not re-run.
🤖 Generated with Claude Code
ppuzio changed pull request status to closed
ppuzio deleted the
refs/pr/96 ref