README.md

---
license: other
license_name: per-source
language:
- pl
task_categories:
- text-generation
pretty_name: GoLLeM Corpus v4 PL (16B tokens)
size_categories:
- 10M<n<100M
configs:
- config_name: default
  data_files: "data/default/*.parquet"
- config_name: cc-by-sa
  data_files: "data/cc-by-sa/*.parquet"
- config_name: web
  data_files: "data/web/*.parquet"
tags:
- pretraining
- polish
- gollem
- from-scratch
- pii-scrubbed
- deduplicated
---

# GoLLeM Corpus v4 PL — SlayerLab/gollem-corpus-16b-pl

**The exact pretraining corpus of the Polish base model GoLLeM v4 (250M, trained from
scratch)** — released *before* training, so the published bytes are byte-identical
(sha-tied) to what the model will see. Successor of
[`SlayerLab/gollem-corpus-2b-pl`](https://huggingface.co/datasets/SlayerLab/gollem-corpus-2b-pl)
(the v2/v3 corpus), scaled ~7.5x with per-record provenance this time.

- **16.58B unique tokens** (GoLLeM V32k tokenizer, measured) =
  ~1.96B curated + 14.62B cleaned Polish web.
- **Per-record `license`, `is_share_alike`, `register`, `source_id` columns** — no blanket license.
- **Three partitions** (Dynaword-style): `default` (permissive), `cc-by-sa` (share-alike
  sources), `web` (HPLT v3 PL, TDM basis — see [License](#license)).
- PII-scrubbed (two-stage pipeline + independent gate), deduplicated (license-aware
  MinHash+LSH), decontaminated against our evaluation sets.

## Composition

Register sizes (V32k tokens, measured this build; sha256 = train `.bin` per register, full set in
[`_manifest.json`](./_manifest.json)):

| register | V32k tokens | docs | train-bin sha256 | main sources |
|---|---:|---:|---|---|
| encyclopedic | 984,163,312 | 1,707,741 | `98a3da0e979c` | Wikipedia PL, Wikisource (CC-BY-SA) |
| science | 522,045,834 | 41,484 | `a5449de3307f` | Biblioteka Nauki (per-record license) |
| news | 210,859,436 | 355,740 | `cf156a47cc20` | polish_news, elka, Wikinews, Global Voices |
| legal/gov | 176,224,723 | 99,152 | `25a5fe007936` | Dziennik Ustaw, gov.pl (PD; 15% register cap) |
| literary | 60,405,659 | 4,228 | `b91f051fc579` | Wolne Lektury, ELTeC-pol (PD/CC-BY-SA) |
| mixed | 1,621,359 | 18,348 | `992d76483f8d` | NKJP-1M |
| **web** | **14,624,198,792** | 17,092,459 | `5045b38564b5` | HPLT v3 PL, cleaned + scrubbed |
| **total** | **16,579,519,115** | 19,319,152 | | train 16,495,148,611 + heldout 84,467,087 |

Held-out (val split, NEVER trained): 84,467,087 V32k tokens / 96,583 docs (sha256 `c4b4990fef60`,
deterministic `sha1(document_id) % 200`). 18 parquet shards across 3 partitions.

Design decisions (documented in the build experiment, DS25B-EXP1):
- **QA sources excluded from base:** PoQuAD is a published Polish QA benchmark — its
  documents in pretraining data would leak evaluation (held out as eval-only). PolQA is
  deferred to the SFT stage (Q/A format skews base-model style).
- **1000 Novels excluded** — residual-author-rights concern despite the CC-BY label
  (mirrors upstream Dynaword's own handling).
- **Legal register capped at 15%** — the dominant failure mode of open Polish corpora is
  legalese skew (71.8% in polish-dynaword v0.2.1); this corpus caps it by construction.
- **Token counts** are reported in the GoLLeM V32k tokenizer (vocab 32,000). Document
  counts are tokenizer-agnostic; re-tokenize with your own tokenizer as needed.

## Quality benchmarks (this corpus)

All numbers from **full-corpus scans of the published artifact** (not samples), pipeline scripts in the build repo:

| check | method | result |
|---|---|---|
| exact duplicates | blake2b/sha8, normalized | 575 docs (0.003%) removed |
| near duplicates | MinHash K=32, 8 bands, word 5-grams (J≈0.6) | 4.80% removed (974,277 of 20,296,635 docs; total removal incl. exact+decontam 4.82%, full-web scan; a 15k sample had estimated 0.03–0.14% — sampling *underestimates* long-tail web duplication, which is why we only report full scans) |
| eval decontamination | exact + word 8-gram fingerprints vs d3_eval, llmzszl, MultiBLiMP, eval_admission | 2,632 docs (0.013%) removed |
| retention | | 95.18% kept (19,319,152 / 20,296,635; falsification threshold pre-registered at >15% removal — not met) |
| license-aware dedup | duplicate survivor chosen by permissiveness (CC0 > CC-BY > CC-BY-SA) | implemented; effect small here (cross-license dups pre-excluded), method contributed for general corpora |
| PII residual (post-gate) | independent full scan, `check ⊋ scrub` | 0.0018% (≤351 / 19,319,152 docs; sha-tie 8/8 verified) |
| trainability smoke | 300 docs/register schema+content audit | schema OK, 0 empty docs, scrub markers present, registers consistent |

## ⚠️ Known benchmark contamination (post-hoc finding — LLMzSzŁ)

**Discovered after training, reported transparently (per our honesty norm).** The build-time
eval-decontamination above (exact + word-8gram vs `llmzszl` et al., 2,632 docs removed) was
**not sufficient** to fully remove Polish state-exam content. A deeper post-hoc re-scan
(12-word fragment match, Aho-Corasick, full 31 GB corpus) found that **~8.83% of LLMzSzŁ
questions (1,662 / 18,820) still have variants/paraphrases in the corpus** — almost entirely
in `web` (1,653) vs curated (9). The build-time 8-gram filter missed reworded/partial variants
that the longer fragment match catches.

**Recomputable fingerprint (verify by recompute, not re-read):**
- `idx_sha` = `a2e4299364f999133e03989230d8730388d804e433f6a190d60d6b1208e079ce`
- recipe: `sha256(json.dumps(sorted(contaminated_item_idx), separators=(",",":")).encode("utf-8"))` — the canonical sha of the sorted index list, **not** a file hash (robust to metadata/wrapper edits; the sha points at the thing, not its packaging).
- number: `len(idx) = 1662`; `1662 / 18820 = 8.83%`; clean `18820 − 1662 = 17158`. Anyone can re-derive every figure here from `idx_sha` — no metadata re-read.
- method: 12-word contiguous fragment match (Aho-Corasick) vs the full 31 GB corpus. `params: not-preserved (pre-norm)` — the scan script/params were not retained, so the method is **documentational, not re-runnable**; the *number* stays recomputable from `idx_sha`, but re-deriving the index from scratch would need a fresh scan.

**Implication:** any benchmark built on **LLMzSzŁ** — and likely other publicly-published
Polish state exams (matura, egzamin ósmoklasisty, egzaminy zawodowe) — will be **inflated**
for models trained on this corpus, because the model saw ~8.83% of the items. This is inherent to
web-crawl corpora: HPLT inherits public exams from the web.

**Recommendation for downstream users:**
- De-contaminate your PL-exam benchmark against this corpus **before** evaluating — use a
  long-fragment or semantic match, not just exact/8-gram.
- For LLMzSzŁ specifically, a `contaminated_idx` (1,662 item indices) is available from the
  build repo; exclude those items and evaluate on the ~17,158 clean.
- Report results with an explicit note: *"de-contaminated subset, ~X% train-seen items
  excluded, fragment-based (paraphrases not eliminated 100%)."*

**General lesson:** exact + 8-gram decontamination at build time under-catches benchmark
leakage from web crawls; a longer-fragment or semantic re-scan **per target benchmark** is
needed. Curated registers (encyclopedia/science) were clean — the leak is concentrated in web.

## Reference downstream results (predecessor models)

The v4 model trained on this corpus will publish its own benchmarks. For calibration,
models trained on this pipeline's predecessor corpus (gollem-corpus-2b-pl, same source
family and cleaning approach):

- **GoLLeM-110M-PL-v3** (2 epochs / ~2B V32k): PL sentiment polemo2 56.5 acc /
  8tags 37.2 (bench_pl harness; +9.3/+5.7 over an independently built 45M PL model on the
  same harness). Cross-lingual control (lm-eval 0.4.12 zero-shot): BLiMP .546,
  SciQ .620, LAMBADA-EN ppl 10508 — i.e. a genuinely Polish model, near-random on
  English, as expected.
- Honest framing: at 110M/2B tokens these are fluency/completion models, ~random on
  zero-shot classification (8tags .11 vs random .125). v4 (250M / 24B seen tokens,
  DataDecide-style 100:1 overtraining) is the scale-up this corpus exists for.

## PII Statement

### Personal Data / PII Processing

All text passed a two-stage PII pipeline before inclusion:

1. **Scrubbing** (`scrub_pii.py`, Paweł, PR#28 + international-format extension): emails → `[PII]`,
   phone numbers (Polish and international formats, label-anchored and self-labelling `+CC`) → `[Telefon]`,
   PESEL/NIP/REGON/KRS and identity-document numbers (checksum-validated, label-anchored) → `[PII]`,
   bank accounts / IBAN (MOD-97 validated) → `[PII]`. Placeholders are semantic tags, not fake values,
   so no synthetic high-frequency numbers poison the corpus.
2. **Independent gate** (second reviewer, lens strictly wider than the scrubber — `check ⊋ scrub`):
   full-corpus residual scan on the exact artifact being published (sha-tied, `verify == publish`).
   Residual after independent gate on the published artifacts: ≤351 documents with unredacted contact data out of 19,319,152 (**0.0018%**), all triaged as institutional/business contacts (university department listings, ministry switchboards, travel-guide venues) in edge formats (+0-trunk, legacy 8-digit); no private-person data found in any triaged sample. Verified sha-tied: all 8 manifest checksums independently recomputed and matched.

Out of scope by explicit decision (ADR): public-institution donation account numbers, court registry
numbers (KRS as bare references), emergency numbers (112/997/998/999), timestamps, coordinates, ISBNs.

Names of public officials in official/government documents are left intact by design (public-interest
information, consistent with the source datasets' practice).

### Known limitations
- Numbers damaged at the source (truncated 8-digit forms) may survive scrubbing; measured share is
  included in the residual figure above.
- The scrubber is regex-based; free-form descriptions of persons are not redacted. The corpus is
  web/official text — downstream users training generative models should apply their own output filters.

## License

This dataset is a compilation of sources with **per-record license tracking** (`license` and
`is_share_alike` columns). It is NOT distributed under a single blanket license.

| Partition | Sources | License basis |
|---|---|---|
| `default` | Wikisource, Wikinews, Wikivoyage (PD/CC0 parts), gov.pl documents, Dziennik Ustaw, Wolne Lektury (PD), NKJP-1M, PoQuAD, ELTeC-pol, Global Voices, news (per-source) | CC0 / public domain / permissive per-source — see per-record `license` column |
| `cc-by-sa` | Wikipedia (994M tok), Wikibooks, Wikiquote, Biblioteka Nauki (SA parts) | CC-BY-SA-4.0 — share-alike applies to derivatives of the TEXT |
| `web` | HPLT v3 PL (cleaned web crawl) | see below |

**HPLT web partition — legal basis statement (important):** the HPLT project releases its *compilation*
under CC0, but explicitly does not own the underlying texts ("we do not own any of the text"). We therefore
do NOT claim CC0 on the web texts themselves. Our redistribution relies on the text-and-data-mining
framework (Directive (EU) 2019/790, art. 4 — lawful access, opt-out respected at crawl level by HPLT);
training use falls under the same TDM basis. Users redistributing or re-publishing the raw texts are
responsible for their own legal review. This mirrors the approach of the Danish Dynaword project.

Attribution for curated sources is provided per-source in the table above and per-record in the
`provenance` column.

## How to use

```python
from datasets import load_dataset

# permissive partition only
ds = load_dataset("SlayerLab/gollem-corpus-16b-pl", "default", split="train")

# full corpus (all partitions; mind per-partition licenses)
for cfg in ("default", "cc-by-sa", "web"):
    part = load_dataset("SlayerLab/gollem-corpus-16b-pl", cfg, split="train")
```

Columns (curated partitions `default`, `cc-by-sa`): `document_id`, `source_id`, `text`,
`register`, `license`, `is_share_alike`, `provenance`, `n_tokens_v32k`.
The `web` partition carries the same fields **except `n_tokens_v32k`** (its per-record V32k
count is not materialized — web tokens were counted at corpus level; see `_manifest.json`
for the web token total and per-shard sha). Load each partition via its own config
(`load_dataset("SlayerLab/gollem-corpus-16b-pl", "<default|cc-by-sa|web>")`), so the schema
difference never surfaces in a single load.

## Provenance & reproducibility

- Build pipeline (scrub → uniform web subsample → license-aware dedup → decontam →
  assembly) is documented step-by-step with parameters in the build experiment log;
  every shard ships a sha256 in the manifest. Published bytes ARE the verified bytes
  (`verify == publish`, checked post-upload by independent re-download).
- Curated layer provenance per-source: [`SlayerLab/polish-dynaword`](https://huggingface.co/datasets/SlayerLab/polish-dynaword).
- Predecessor corpus (v2/v3, 2B): [`SlayerLab/gollem-corpus-2b-pl`](https://huggingface.co/datasets/SlayerLab/gollem-corpus-2b-pl).
- Design predecessor / scaffold: [`SlayerLab/slayer-pl-8x3b`](https://huggingface.co/datasets/SlayerLab/slayer-pl-8x3b) — same design and source selection (8-pack, ~24B-token plan); this corpus is the **materialized + cleaned** realization of that plan (adds full-scan near-dedup and the PII gate the scaffold did not have).
- Model line: [`SlayerLab/GoLLeM-110M-PL-v3`](https://huggingface.co/SlayerLab/GoLLeM-110M-PL-v3) → GoLLeM v4 (250M, this corpus).

## Acknowledgements

- PII scrubbing: Paweł Puzio ([ppuzio](https://huggingface.co/ppuzio), PR#28 in polish-dynaword)
- Independent PII gate + international-format fix: Wartownik (N-04, Slayer Kolektyw)
- License-aware dedup + per-license partitions: Monter (N-03, Slayer Kolektyw)
- Card integration & training: Hart (N-02, Slayer Kolektyw)

**Author:** Arkadiusz Słota / SlayerLab.

## Podsumowanie (PL)

Dokładny korpus pretreningowy polskiego modelu bazowego **GoLLeM v4 (250M, od zera)** —
publikowany PRZED treningiem, bajt-w-bajt ten sam zbiór (sha-tied), na którym model będzie
uczony. 16.58B unikatowych tokenów V32k: ~1.96B kuratorowanych (encyklopedia, nauka, news,
legal z capem 15%, literatura) + 14.62B oczyszczonego polskiego weba (HPLT v3). Licencje
per-rekord w trzech partycjach (permissive / CC-BY-SA / web-TDM), PII po dwustopniowym
scrubie z niezależną bramką (residual 0.0018%), dedup license-aware z pełnego skanu (retencja 95.18%),
dekontaminacja względem naszych zbiorów ewaluacyjnych.