File size: 4,314 Bytes
d2e00f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
---
license: cc-by-sa-4.0
language:
- pl
task_categories:
- text-generation
pretty_name: GoLLeM Corpus 2B PL
size_categories:
- 1M<n<10M
tags:
- pretraining
- polish
- gollem
- from-scratch
---

# GoLLeM Corpus 2B PL

**Dokładny korpus treningowy** polskiego modelu bazowego
[`SlayerLab/GoLLeM-110M-PL-v3`](https://huggingface.co/SlayerLab/GoLLeM-110M-PL-v3)
(oraz v2) — ten sam zbiór, po którym model przeszedł **dwie epoki**. Publikujemy go,
aby każdy mógł **odtworzyć trening od zera** na własnym tokenizerze.

> **Jak powstał ten plik.** Korpus odzyskano przez **zdekodowanie** stokenizowanego
> checkpointu treningowego (byte-level BPE `dynaword-32k`, round-trip bezstratny; granice
> dokumentów = token `<|endoftext|>`). To jest **dokładnie** tekst, który model widział —
> nie rekonstrukcja ze źródeł.

## Zawartość
| split | dokumenty | tokeny (proxy dynaword-32k) |
|---|---:|---:|
| train | 2 790 164 | 1 987 316 263 |
| validation | 14 021 | 10 123 670 |

Format: Parquet (zstd), kolumny `id` (int), `text` (str), `n_tokens` (int, w tokenizerze
dynaword-32k). Liczba tokenów jest orientacyjna — przy własnym tokenizerze przelicz sam;
liczba **dokumentów** jest tokenizer-agnostyczna.

## Skład (cele, % tokenów)
- **~58% curated** (wiedza + literatura): Wikipedia, Wikisource, Wolne Lektury, 1000 Novels,
  Wikiquote / Wikibooks / Wikinews / Wikivoyage, ELTeC-pol.
- **~42% web**: HPLT v3.0 PL (po czyszczeniu: strip boilerplate span+line, filtr jakości/języka).
- **Zero legalese** — świadomie wykluczone eur-lex / parlament / Dziennik Ustaw, by rejestr
  prawny nie zdominował korpusu.

## Jak używać
```python
from datasets import load_dataset
ds = load_dataset("SlayerLab/gollem-corpus-2b-pl", split="train")
print(ds[0]["text"][:300])
# ... następnie tokenizuj WŁASNYM tokenizerem i trenuj model bazowy
```

## Zastrzeżenia (uczciwie)
- **Surowy korpus pretreningowy**, nie zbiór wyczyszczony pod czytelnika: część web zawiera
  listy słów-kluczy, tagi itp. — to realne dane treningowe (42% HPLT).
- **Proweniencja agregatowa.** Ponieważ korpus odzyskano z jednolitego strumienia tokenów,
  **etykiety źródła/licencji per dokument nie są zachowane** — udokumentowany jest jedynie skład
  zbiorczy (wyżej). Jeśli potrzebujesz proweniencji per-źródło z datasheetami, użyj
  [`SlayerLab/polish-dynaword`](https://huggingface.co/datasets/SlayerLab/polish-dynaword)
  (z niego pochodzi warstwa curated).
- **PII scrubowane** na etapie treningu (telefon/email/PESEL/NIP → tagi) — korpus nie jest źródłem PII.
- Tokeny liczone tokenizerem `dynaword-32k` (proxy), nie liczbą słów/bajtów.

## Licencja i atrybucja
**CC-BY-SA-4.0** (licencja dominująca, share-alike). Skład licencyjny źródeł:
Wikimedia (Wikipedia/Wikisource/Wiki*) — **CC-BY-SA-3.0**; Wolne Lektury — **CC-BY-SA-4.0 /
Wolna Sztuka 1.3**; 1000 Novels, ELTeC-pol — **CC-BY-4.0**; HPLT v3 (web) — **CC0-1.0**.
Użycie wymaga **atrybucji** (Wikimedia Foundation, Wolne Lektury, autorzy 1000 Novels,
HPLT / CLARIN-PL) oraz **share-alike** dla derywatów.

## Powiązania
- Model trenowany na tym korpusie: [`SlayerLab/GoLLeM-110M-PL-v3`](https://huggingface.co/SlayerLab/GoLLeM-110M-PL-v3)
- Warstwa curated / proweniencja per-źródło: [`SlayerLab/polish-dynaword`](https://huggingface.co/datasets/SlayerLab/polish-dynaword)

**Autor:** Arkadiusz Słota / SlayerLab.

---

## English (summary)

Exact pretraining corpus of **GoLLeM-110M-PL v2/v3** — the same data the model saw for two
epochs — released so anyone can **retrain from scratch** with their own tokenizer. Recovered by
**decoding** the tokenized training checkpoint (byte-level BPE, lossless round-trip; documents
split on `<|endoftext|>`), so it is *exactly* what the model was trained on.

**train** 2,790,164 docs / 1.99B tokens (dynaword-32k proxy); **validation** 14,021 docs.
Mix: ~58% curated (Wikimedia, Wolne Lektury, 1000 Novels, ELTeC) + ~42% cleaned HPLT v3 PL web;
**no legalese**. Raw pretraining data (web keyword-lists included). **Aggregate provenance only**
(per-document source/license labels are not preserved by the decode; for per-source datasheets use
`SlayerLab/polish-dynaword`). PII scrubbed during training. **CC-BY-SA-4.0**, attribution +
share-alike required.