Add Known benchmark contamination section (LLMzSzL ~9% post-hoc finding, decontam-first); drop invalid license_link anchor
Browse files
README.md
CHANGED
|
@@ -1,201 +1,226 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: other
|
| 3 |
-
license_name: per-source
|
| 4 |
-
language:
|
| 5 |
-
- pl
|
| 6 |
-
task_categories:
|
| 7 |
-
- text-generation
|
| 8 |
-
pretty_name: GoLLeM Corpus v4 PL (16B tokens)
|
| 9 |
-
size_categories:
|
| 10 |
-
- 10M<n<100M
|
| 11 |
-
configs:
|
| 12 |
-
- config_name: default
|
| 13 |
-
data_files: "data/default/*.parquet"
|
| 14 |
-
- config_name: cc-by-sa
|
| 15 |
-
data_files: "data/cc-by-sa/*.parquet"
|
| 16 |
-
- config_name: web
|
| 17 |
-
data_files: "data/web/*.parquet"
|
| 18 |
-
tags:
|
| 19 |
-
- pretraining
|
| 20 |
-
- polish
|
| 21 |
-
- gollem
|
| 22 |
-
- from-scratch
|
| 23 |
-
- pii-scrubbed
|
| 24 |
-
- deduplicated
|
| 25 |
-
---
|
| 26 |
-
|
| 27 |
-
# GoLLeM Corpus v4 PL — SlayerLab/gollem-corpus-16b-pl
|
| 28 |
-
|
| 29 |
-
**The exact pretraining corpus of the Polish base model GoLLeM v4 (250M, trained from
|
| 30 |
-
scratch)** — released *before* training, so the published bytes are byte-identical
|
| 31 |
-
(sha-tied) to what the model will see. Successor of
|
| 32 |
-
[`SlayerLab/gollem-corpus-2b-pl`](https://huggingface.co/datasets/SlayerLab/gollem-corpus-2b-pl)
|
| 33 |
-
(the v2/v3 corpus), scaled ~7.5x with per-record provenance this time.
|
| 34 |
-
|
| 35 |
-
- **16.58B unique tokens** (GoLLeM V32k tokenizer, measured) =
|
| 36 |
-
~1.96B curated + 14.62B cleaned Polish web.
|
| 37 |
-
- **Per-record `license`, `is_share_alike`, `register`, `source_id` columns** — no blanket license.
|
| 38 |
-
- **Three partitions** (Dynaword-style): `default` (permissive), `cc-by-sa` (share-alike
|
| 39 |
-
sources), `web` (HPLT v3 PL, TDM basis — see [License](#license)).
|
| 40 |
-
- PII-scrubbed (two-stage pipeline + independent gate), deduplicated (license-aware
|
| 41 |
-
MinHash+LSH), decontaminated against our evaluation sets.
|
| 42 |
-
|
| 43 |
-
## Composition
|
| 44 |
-
|
| 45 |
-
Register sizes (V32k tokens, measured this build; sha256 = train `.bin` per register, full set in
|
| 46 |
-
[`_manifest.json`](./_manifest.json)):
|
| 47 |
-
|
| 48 |
-
| register | V32k tokens | docs | train-bin sha256 | main sources |
|
| 49 |
-
|---|---:|---:|---|---|
|
| 50 |
-
| encyclopedic | 984,163,312 | 1,707,741 | `98a3da0e979c` | Wikipedia PL, Wikisource (CC-BY-SA) |
|
| 51 |
-
| science | 522,045,834 | 41,484 | `a5449de3307f` | Biblioteka Nauki (per-record license) |
|
| 52 |
-
| news | 210,859,436 | 355,740 | `cf156a47cc20` | polish_news, elka, Wikinews, Global Voices |
|
| 53 |
-
| legal/gov | 176,224,723 | 99,152 | `25a5fe007936` | Dziennik Ustaw, gov.pl (PD; 15% register cap) |
|
| 54 |
-
| literary | 60,405,659 | 4,228 | `b91f051fc579` | Wolne Lektury, ELTeC-pol (PD/CC-BY-SA) |
|
| 55 |
-
| mixed | 1,621,359 | 18,348 | `992d76483f8d` | NKJP-1M |
|
| 56 |
-
| **web** | **14,624,198,792** | 17,092,459 | `5045b38564b5` | HPLT v3 PL, cleaned + scrubbed |
|
| 57 |
-
| **total** | **16,579,519,115** | 19,319,152 | | train 16,495,148,611 + heldout 84,467,087 |
|
| 58 |
-
|
| 59 |
-
Held-out (val split, NEVER trained): 84,467,087 V32k tokens / 96,583 docs (sha256 `c4b4990fef60`,
|
| 60 |
-
deterministic `sha1(document_id) % 200`). 18 parquet shards across 3 partitions.
|
| 61 |
-
|
| 62 |
-
Design decisions (documented in the build experiment, DS25B-EXP1):
|
| 63 |
-
- **QA sources excluded from base:** PoQuAD is a published Polish QA benchmark — its
|
| 64 |
-
documents in pretraining data would leak evaluation (held out as eval-only). PolQA is
|
| 65 |
-
deferred to the SFT stage (Q/A format skews base-model style).
|
| 66 |
-
- **1000 Novels excluded** — residual-author-rights concern despite the CC-BY label
|
| 67 |
-
(mirrors upstream Dynaword's own handling).
|
| 68 |
-
- **Legal register capped at 15%** — the dominant failure mode of open Polish corpora is
|
| 69 |
-
legalese skew (71.8% in polish-dynaword v0.2.1); this corpus caps it by construction.
|
| 70 |
-
- **Token counts** are reported in the GoLLeM V32k tokenizer (vocab 32,000). Document
|
| 71 |
-
counts are tokenizer-agnostic; re-tokenize with your own tokenizer as needed.
|
| 72 |
-
|
| 73 |
-
## Quality benchmarks (this corpus)
|
| 74 |
-
|
| 75 |
-
All numbers from **full-corpus scans of the published artifact** (not samples), pipeline scripts in the build repo:
|
| 76 |
-
|
| 77 |
-
| check | method | result |
|
| 78 |
-
|---|---|---|
|
| 79 |
-
| exact duplicates | blake2b/sha8, normalized | 575 docs (0.003%) removed |
|
| 80 |
-
| 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) |
|
| 81 |
-
| eval decontamination | exact + word 8-gram fingerprints vs d3_eval, llmzszl, MultiBLiMP, eval_admission | 2,632 docs (0.013%) removed |
|
| 82 |
-
| retention | | 95.18% kept (19,319,152 / 20,296,635; falsification threshold pre-registered at >15% removal — not met) |
|
| 83 |
-
| 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 |
|
| 84 |
-
| PII residual (post-gate) | independent full scan, `check ⊋ scrub` | 0.0018% (≤351 / 19,319,152 docs; sha-tie 8/8 verified) |
|
| 85 |
-
| trainability smoke | 300 docs/register schema+content audit | schema OK, 0 empty docs, scrub markers present, registers consistent |
|
| 86 |
-
|
| 87 |
-
##
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
-
|
| 126 |
-
|
| 127 |
-
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
```
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
(
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: per-source
|
| 4 |
+
language:
|
| 5 |
+
- pl
|
| 6 |
+
task_categories:
|
| 7 |
+
- text-generation
|
| 8 |
+
pretty_name: GoLLeM Corpus v4 PL (16B tokens)
|
| 9 |
+
size_categories:
|
| 10 |
+
- 10M<n<100M
|
| 11 |
+
configs:
|
| 12 |
+
- config_name: default
|
| 13 |
+
data_files: "data/default/*.parquet"
|
| 14 |
+
- config_name: cc-by-sa
|
| 15 |
+
data_files: "data/cc-by-sa/*.parquet"
|
| 16 |
+
- config_name: web
|
| 17 |
+
data_files: "data/web/*.parquet"
|
| 18 |
+
tags:
|
| 19 |
+
- pretraining
|
| 20 |
+
- polish
|
| 21 |
+
- gollem
|
| 22 |
+
- from-scratch
|
| 23 |
+
- pii-scrubbed
|
| 24 |
+
- deduplicated
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
# GoLLeM Corpus v4 PL — SlayerLab/gollem-corpus-16b-pl
|
| 28 |
+
|
| 29 |
+
**The exact pretraining corpus of the Polish base model GoLLeM v4 (250M, trained from
|
| 30 |
+
scratch)** — released *before* training, so the published bytes are byte-identical
|
| 31 |
+
(sha-tied) to what the model will see. Successor of
|
| 32 |
+
[`SlayerLab/gollem-corpus-2b-pl`](https://huggingface.co/datasets/SlayerLab/gollem-corpus-2b-pl)
|
| 33 |
+
(the v2/v3 corpus), scaled ~7.5x with per-record provenance this time.
|
| 34 |
+
|
| 35 |
+
- **16.58B unique tokens** (GoLLeM V32k tokenizer, measured) =
|
| 36 |
+
~1.96B curated + 14.62B cleaned Polish web.
|
| 37 |
+
- **Per-record `license`, `is_share_alike`, `register`, `source_id` columns** — no blanket license.
|
| 38 |
+
- **Three partitions** (Dynaword-style): `default` (permissive), `cc-by-sa` (share-alike
|
| 39 |
+
sources), `web` (HPLT v3 PL, TDM basis — see [License](#license)).
|
| 40 |
+
- PII-scrubbed (two-stage pipeline + independent gate), deduplicated (license-aware
|
| 41 |
+
MinHash+LSH), decontaminated against our evaluation sets.
|
| 42 |
+
|
| 43 |
+
## Composition
|
| 44 |
+
|
| 45 |
+
Register sizes (V32k tokens, measured this build; sha256 = train `.bin` per register, full set in
|
| 46 |
+
[`_manifest.json`](./_manifest.json)):
|
| 47 |
+
|
| 48 |
+
| register | V32k tokens | docs | train-bin sha256 | main sources |
|
| 49 |
+
|---|---:|---:|---|---|
|
| 50 |
+
| encyclopedic | 984,163,312 | 1,707,741 | `98a3da0e979c` | Wikipedia PL, Wikisource (CC-BY-SA) |
|
| 51 |
+
| science | 522,045,834 | 41,484 | `a5449de3307f` | Biblioteka Nauki (per-record license) |
|
| 52 |
+
| news | 210,859,436 | 355,740 | `cf156a47cc20` | polish_news, elka, Wikinews, Global Voices |
|
| 53 |
+
| legal/gov | 176,224,723 | 99,152 | `25a5fe007936` | Dziennik Ustaw, gov.pl (PD; 15% register cap) |
|
| 54 |
+
| literary | 60,405,659 | 4,228 | `b91f051fc579` | Wolne Lektury, ELTeC-pol (PD/CC-BY-SA) |
|
| 55 |
+
| mixed | 1,621,359 | 18,348 | `992d76483f8d` | NKJP-1M |
|
| 56 |
+
| **web** | **14,624,198,792** | 17,092,459 | `5045b38564b5` | HPLT v3 PL, cleaned + scrubbed |
|
| 57 |
+
| **total** | **16,579,519,115** | 19,319,152 | | train 16,495,148,611 + heldout 84,467,087 |
|
| 58 |
+
|
| 59 |
+
Held-out (val split, NEVER trained): 84,467,087 V32k tokens / 96,583 docs (sha256 `c4b4990fef60`,
|
| 60 |
+
deterministic `sha1(document_id) % 200`). 18 parquet shards across 3 partitions.
|
| 61 |
+
|
| 62 |
+
Design decisions (documented in the build experiment, DS25B-EXP1):
|
| 63 |
+
- **QA sources excluded from base:** PoQuAD is a published Polish QA benchmark — its
|
| 64 |
+
documents in pretraining data would leak evaluation (held out as eval-only). PolQA is
|
| 65 |
+
deferred to the SFT stage (Q/A format skews base-model style).
|
| 66 |
+
- **1000 Novels excluded** — residual-author-rights concern despite the CC-BY label
|
| 67 |
+
(mirrors upstream Dynaword's own handling).
|
| 68 |
+
- **Legal register capped at 15%** — the dominant failure mode of open Polish corpora is
|
| 69 |
+
legalese skew (71.8% in polish-dynaword v0.2.1); this corpus caps it by construction.
|
| 70 |
+
- **Token counts** are reported in the GoLLeM V32k tokenizer (vocab 32,000). Document
|
| 71 |
+
counts are tokenizer-agnostic; re-tokenize with your own tokenizer as needed.
|
| 72 |
+
|
| 73 |
+
## Quality benchmarks (this corpus)
|
| 74 |
+
|
| 75 |
+
All numbers from **full-corpus scans of the published artifact** (not samples), pipeline scripts in the build repo:
|
| 76 |
+
|
| 77 |
+
| check | method | result |
|
| 78 |
+
|---|---|---|
|
| 79 |
+
| exact duplicates | blake2b/sha8, normalized | 575 docs (0.003%) removed |
|
| 80 |
+
| 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) |
|
| 81 |
+
| eval decontamination | exact + word 8-gram fingerprints vs d3_eval, llmzszl, MultiBLiMP, eval_admission | 2,632 docs (0.013%) removed |
|
| 82 |
+
| retention | | 95.18% kept (19,319,152 / 20,296,635; falsification threshold pre-registered at >15% removal — not met) |
|
| 83 |
+
| 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 |
|
| 84 |
+
| PII residual (post-gate) | independent full scan, `check ⊋ scrub` | 0.0018% (≤351 / 19,319,152 docs; sha-tie 8/8 verified) |
|
| 85 |
+
| trainability smoke | 300 docs/register schema+content audit | schema OK, 0 empty docs, scrub markers present, registers consistent |
|
| 86 |
+
|
| 87 |
+
## ⚠️ Known benchmark contamination (post-hoc finding — LLMzSzŁ)
|
| 88 |
+
|
| 89 |
+
**Discovered after training, reported transparently (per our honesty norm).** The build-time
|
| 90 |
+
eval-decontamination above (exact + word-8gram vs `llmzszl` et al., 2,632 docs removed) was
|
| 91 |
+
**not sufficient** to fully remove Polish state-exam content. A deeper post-hoc re-scan
|
| 92 |
+
(12-word fragment match, Aho-Corasick, full 31 GB corpus) found that **~9% of LLMzSzŁ
|
| 93 |
+
questions (1,662 / 18,430) still have variants/paraphrases in the corpus** — almost entirely
|
| 94 |
+
in `web` (1,653) vs curated (9). The build-time 8-gram filter missed reworded/partial variants
|
| 95 |
+
that the longer fragment match catches.
|
| 96 |
+
|
| 97 |
+
**Implication:** any benchmark built on **LLMzSzŁ** — and likely other publicly-published
|
| 98 |
+
Polish state exams (matura, egzamin ósmoklasisty, egzaminy zawodowe) — will be **inflated**
|
| 99 |
+
for models trained on this corpus, because the model saw ~9% of the items. This is inherent to
|
| 100 |
+
web-crawl corpora: HPLT inherits public exams from the web.
|
| 101 |
+
|
| 102 |
+
**Recommendation for downstream users:**
|
| 103 |
+
- De-contaminate your PL-exam benchmark against this corpus **before** evaluating — use a
|
| 104 |
+
long-fragment or semantic match, not just exact/8-gram.
|
| 105 |
+
- For LLMzSzŁ specifically, a `contaminated_idx` (1,662 item indices) is available from the
|
| 106 |
+
build repo; exclude those items and evaluate on the ~16,768 clean.
|
| 107 |
+
- Report results with an explicit note: *"de-contaminated subset, ~X% train-seen items
|
| 108 |
+
excluded, fragment-based (paraphrases not eliminated 100%)."*
|
| 109 |
+
|
| 110 |
+
**General lesson:** exact + 8-gram decontamination at build time under-catches benchmark
|
| 111 |
+
leakage from web crawls; a longer-fragment or semantic re-scan **per target benchmark** is
|
| 112 |
+
needed. Curated registers (encyclopedia/science) were clean — the leak is concentrated in web.
|
| 113 |
+
|
| 114 |
+
## Reference downstream results (predecessor models)
|
| 115 |
+
|
| 116 |
+
The v4 model trained on this corpus will publish its own benchmarks. For calibration,
|
| 117 |
+
models trained on this pipeline's predecessor corpus (gollem-corpus-2b-pl, same source
|
| 118 |
+
family and cleaning approach):
|
| 119 |
+
|
| 120 |
+
- **GoLLeM-110M-PL-v3** (2 epochs / ~2B V32k): PL sentiment polemo2 56.5 acc /
|
| 121 |
+
8tags 37.2 (bench_pl harness; +9.3/+5.7 over an independently built 45M PL model on the
|
| 122 |
+
same harness). Cross-lingual control (lm-eval 0.4.12 zero-shot): BLiMP .546,
|
| 123 |
+
SciQ .620, LAMBADA-EN ppl 10508 — i.e. a genuinely Polish model, near-random on
|
| 124 |
+
English, as expected.
|
| 125 |
+
- Honest framing: at 110M/2B tokens these are fluency/completion models, ~random on
|
| 126 |
+
zero-shot classification (8tags .11 vs random .125). v4 (250M / 24B seen tokens,
|
| 127 |
+
DataDecide-style 100:1 overtraining) is the scale-up this corpus exists for.
|
| 128 |
+
|
| 129 |
+
## PII Statement
|
| 130 |
+
|
| 131 |
+
### Personal Data / PII Processing
|
| 132 |
+
|
| 133 |
+
All text passed a two-stage PII pipeline before inclusion:
|
| 134 |
+
|
| 135 |
+
1. **Scrubbing** (`scrub_pii.py`, Paweł, PR#28 + international-format extension): emails → `[PII]`,
|
| 136 |
+
phone numbers (Polish and international formats, label-anchored and self-labelling `+CC`) → `[Telefon]`,
|
| 137 |
+
PESEL/NIP/REGON/KRS and identity-document numbers (checksum-validated, label-anchored) → `[PII]`,
|
| 138 |
+
bank accounts / IBAN (MOD-97 validated) → `[PII]`. Placeholders are semantic tags, not fake values,
|
| 139 |
+
so no synthetic high-frequency numbers poison the corpus.
|
| 140 |
+
2. **Independent gate** (second reviewer, lens strictly wider than the scrubber — `check ⊋ scrub`):
|
| 141 |
+
full-corpus residual scan on the exact artifact being published (sha-tied, `verify == publish`).
|
| 142 |
+
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.
|
| 143 |
+
|
| 144 |
+
Out of scope by explicit decision (ADR): public-institution donation account numbers, court registry
|
| 145 |
+
numbers (KRS as bare references), emergency numbers (112/997/998/999), timestamps, coordinates, ISBNs.
|
| 146 |
+
|
| 147 |
+
Names of public officials in official/government documents are left intact by design (public-interest
|
| 148 |
+
information, consistent with the source datasets' practice).
|
| 149 |
+
|
| 150 |
+
### Known limitations
|
| 151 |
+
- Numbers damaged at the source (truncated 8-digit forms) may survive scrubbing; measured share is
|
| 152 |
+
included in the residual figure above.
|
| 153 |
+
- The scrubber is regex-based; free-form descriptions of persons are not redacted. The corpus is
|
| 154 |
+
web/official text — downstream users training generative models should apply their own output filters.
|
| 155 |
+
|
| 156 |
+
## License
|
| 157 |
+
|
| 158 |
+
This dataset is a compilation of sources with **per-record license tracking** (`license` and
|
| 159 |
+
`is_share_alike` columns). It is NOT distributed under a single blanket license.
|
| 160 |
+
|
| 161 |
+
| Partition | Sources | License basis |
|
| 162 |
+
|---|---|---|
|
| 163 |
+
| `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 |
|
| 164 |
+
| `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 |
|
| 165 |
+
| `web` | HPLT v3 PL (cleaned web crawl) | see below |
|
| 166 |
+
|
| 167 |
+
**HPLT web partition — legal basis statement (important):** the HPLT project releases its *compilation*
|
| 168 |
+
under CC0, but explicitly does not own the underlying texts ("we do not own any of the text"). We therefore
|
| 169 |
+
do NOT claim CC0 on the web texts themselves. Our redistribution relies on the text-and-data-mining
|
| 170 |
+
framework (Directive (EU) 2019/790, art. 4 — lawful access, opt-out respected at crawl level by HPLT);
|
| 171 |
+
training use falls under the same TDM basis. Users redistributing or re-publishing the raw texts are
|
| 172 |
+
responsible for their own legal review. This mirrors the approach of the Danish Dynaword project.
|
| 173 |
+
|
| 174 |
+
Attribution for curated sources is provided per-source in the table above and per-record in the
|
| 175 |
+
`provenance` column.
|
| 176 |
+
|
| 177 |
+
## How to use
|
| 178 |
+
|
| 179 |
+
```python
|
| 180 |
+
from datasets import load_dataset
|
| 181 |
+
|
| 182 |
+
# permissive partition only
|
| 183 |
+
ds = load_dataset("SlayerLab/gollem-corpus-16b-pl", "default", split="train")
|
| 184 |
+
|
| 185 |
+
# full corpus (all partitions; mind per-partition licenses)
|
| 186 |
+
for cfg in ("default", "cc-by-sa", "web"):
|
| 187 |
+
part = load_dataset("SlayerLab/gollem-corpus-16b-pl", cfg, split="train")
|
| 188 |
+
```
|
| 189 |
+
|
| 190 |
+
Columns (curated partitions `default`, `cc-by-sa`): `document_id`, `source_id`, `text`,
|
| 191 |
+
`register`, `license`, `is_share_alike`, `provenance`, `n_tokens_v32k`.
|
| 192 |
+
The `web` partition carries the same fields **except `n_tokens_v32k`** (its per-record V32k
|
| 193 |
+
count is not materialized — web tokens were counted at corpus level; see `_manifest.json`
|
| 194 |
+
for the web token total and per-shard sha). Load each partition via its own config
|
| 195 |
+
(`load_dataset("SlayerLab/gollem-corpus-16b-pl", "<default|cc-by-sa|web>")`), so the schema
|
| 196 |
+
difference never surfaces in a single load.
|
| 197 |
+
|
| 198 |
+
## Provenance & reproducibility
|
| 199 |
+
|
| 200 |
+
- Build pipeline (scrub → uniform web subsample → license-aware dedup → decontam →
|
| 201 |
+
assembly) is documented step-by-step with parameters in the build experiment log;
|
| 202 |
+
every shard ships a sha256 in the manifest. Published bytes ARE the verified bytes
|
| 203 |
+
(`verify == publish`, checked post-upload by independent re-download).
|
| 204 |
+
- Curated layer provenance per-source: [`SlayerLab/polish-dynaword`](https://huggingface.co/datasets/SlayerLab/polish-dynaword).
|
| 205 |
+
- Predecessor corpus (v2/v3, 2B): [`SlayerLab/gollem-corpus-2b-pl`](https://huggingface.co/datasets/SlayerLab/gollem-corpus-2b-pl).
|
| 206 |
+
- 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).
|
| 207 |
+
- Model line: [`SlayerLab/GoLLeM-110M-PL-v3`](https://huggingface.co/SlayerLab/GoLLeM-110M-PL-v3) → GoLLeM v4 (250M, this corpus).
|
| 208 |
+
|
| 209 |
+
## Acknowledgements
|
| 210 |
+
|
| 211 |
+
- PII scrubbing: Paweł Puzio ([ppuzio](https://huggingface.co/ppuzio), PR#28 in polish-dynaword)
|
| 212 |
+
- Independent PII gate + international-format fix: Wartownik (N-04, Slayer Kolektyw)
|
| 213 |
+
- License-aware dedup + per-license partitions: Monter (N-03, Slayer Kolektyw)
|
| 214 |
+
- Card integration & training: Hart (N-02, Slayer Kolektyw)
|
| 215 |
+
|
| 216 |
+
**Author:** Arkadiusz Słota / SlayerLab.
|
| 217 |
+
|
| 218 |
+
## Podsumowanie (PL)
|
| 219 |
+
|
| 220 |
+
Dokładny korpus pretreningowy polskiego modelu bazowego **GoLLeM v4 (250M, od zera)** —
|
| 221 |
+
publikowany PRZED treningiem, bajt-w-bajt ten sam zbiór (sha-tied), na którym model będzie
|
| 222 |
+
uczony. 16.58B unikatowych tokenów V32k: ~1.96B kuratorowanych (encyklopedia, nauka, news,
|
| 223 |
+
legal z capem 15%, literatura) + 14.62B oczyszczonego polskiego weba (HPLT v3). Licencje
|
| 224 |
+
per-rekord w trzech partycjach (permissive / CC-BY-SA / web-TDM), PII po dwustopniowym
|
| 225 |
+
scrubie z niezależną bramką (residual 0.0018%), dedup license-aware z pełnego skanu (retencja 95.18%),
|
| 226 |
+
dekontaminacja względem naszych zbiorów ewaluacyjnych.
|