File size: 5,136 Bytes
efdf4e6 | 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 94 95 96 97 | ---
license: other
license_name: per-source
license_link: https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md
language:
- en
- fr
pretty_name: skycmd pretraining text and tokenizers
task_categories:
- fill-mask
tags:
- skycmd
configs:
- config_name: dolma3_finemath
data_files: data/dolma3_finemath.parquet
- config_name: dolma3_science_pdfs
data_files: data/dolma3_science_pdfs.parquet
- config_name: dolma3_wiki_en
data_files: data/dolma3_wiki_en.parquet
- config_name: fineweb2_fr
data_files: data/fineweb2_fr.parquet
- config_name: fineweb_edu
data_files: data/fineweb_edu.parquet
- config_name: simplestories
data_files: data/simplestories.parquet
- config_name: tinystories
data_files: data/tinystories.parquet
- config_name: tinystories_fr
data_files: data/tinystories_fr.parquet
- config_name: txt360_europarl_fr
data_files: data/txt360_europarl_fr.parquet
- config_name: txt360_pg19
data_files: data/txt360_pg19.parquet
- config_name: txt360_stackexchange
data_files: data/txt360_stackexchange.parquet
---
# skycmd pretraining text and tokenizers
The text of skycmd's masked-language-model pretraining (s1 to s6), English and French, and the tokenizers trained on it. `data/text` is read by every model; `data/text-open` (non-web parts of Dolma 3 and TxT360) is read first by the s6 models.
Part of [skycmd](https://github.com/maxgfr/skycmd), tiny decision models trained from scratch in Rust that run in the
browser. Written by `sky-data fetch-text and sky-data fetch-text --set open (scripts/prepare-data.sh)` at commit [`84c6d854`](https://github.com/maxgfr/skycmd/commit/84c6d854cdc2a9f6d951ebb55780060a01f38ca6).
Licences of every set: [DATA_LICENSES.md](https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md).
Each row is one document: `src` (the source) and `text`. All of it is deduplicated exactly and with
MinHash/LSH (word 5-grams, Jaccard >= 0.8). The tokenizers trained on it are in `tokenizers/`
(`sky-data train-tokenizer --vocab N --num`).
## Sources and licences
Each config is one source and keeps **its source's licence** (share-alike sources stay share-alike).
Attribution goes to the upstream authors; the conversion is ours.
| config | upstream | licence | how it is used | rows |
|---|---|---|---|---:|
| `dolma3_finemath` | allenai/dolma3_mix-150B-1025 `data/finemath-3plus` | odc-by (FineMath, Dolma 3) and CommonCrawl's terms of use | shard `00000000`, exact duplicates dropped | 4,354 |
| `dolma3_science_pdfs` | allenai/dolma3_mix-150B-1025 `data/olmocr_science_pdfs-science_math_and_technology` | odc-by (Dolma 3); the PDFs keep their authors' rights, see the Dolma 3 card | shard `00000000`, olmOCR transcriptions | 1,517 |
| `dolma3_wiki_en` | allenai/dolma3_mix-150B-1025 `data/dolma1_7-wiki-en` | odc-by (Dolma 3); Wikipedia and Wikibooks text CC-BY-SA-4.0 / GFDL | shard `00000000` | 2,489 |
| `fineweb2_fr` | HuggingFaceFW/fineweb-2 (`fra_Latn`, shard `train/000_00000`) | odc-by (and CommonCrawl's terms of use) | first row groups, 34 MB | 11,010 |
| `fineweb_edu` | HuggingFaceFW/fineweb-edu (`sample/10BT`, shard `000_00000`) | odc-by (and CommonCrawl's terms of use) | first row groups, 37.5 MB | 9,002 |
| `simplestories` | SimpleStories/SimpleStories | mit | `story` column, first row groups of train, 30 MB | 24,248 |
| `tinystories` | roneneldan/TinyStories | cdla-sharing-1.0 | `text` column, first row groups of train, 30 MB | 33,400 |
| `tinystories_fr` | btecsec/tinystories-fr | cdla-sharing-1.0 | French `text`, whole train split | 20,000 |
| `txt360_europarl_fr` | IFM/TxT360 `data/europarl` (French rows) | odc-by (TxT360); European Parliament proceedings, reuse authorised with acknowledgement | up to 8 pieces per session, 15 MB | 873 |
| `txt360_pg19` | IFM/TxT360 `data/pg19` | odc-by (TxT360); the books are public domain (Project Gutenberg) | up to 8 pieces of 20,000 characters per book | 1,140 |
| `txt360_stackexchange` | IFM/TxT360 `data/stackexchange` | odc-by (TxT360); Stack Exchange posts CC-BY-SA | one shard, 15 MB | 4,638 |
## Files
| file | rows |
|---|---:|
| `data/dolma3_finemath.parquet` | 4,354 |
| `data/dolma3_science_pdfs.parquet` | 1,517 |
| `data/dolma3_wiki_en.parquet` | 2,489 |
| `data/fineweb2_fr.parquet` | 11,010 |
| `data/fineweb_edu.parquet` | 9,002 |
| `data/simplestories.parquet` | 24,248 |
| `data/tinystories.parquet` | 33,400 |
| `data/tinystories_fr.parquet` | 20,000 |
| `data/txt360_europarl_fr.parquet` | 873 |
| `data/txt360_pg19.parquet` | 1,140 |
| `data/txt360_stackexchange.parquet` | 4,638 |
| `tokenizers/tok-s2-1024.json` | — |
| `tokenizers/tok-s6-2048.json` | — |
| `tokenizers/tok-s6-4096.json` | — |
```python
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-text", "dolma3_finemath")
```
skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream
sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.
TinyStories, TinyStories-fr and SimpleStories are synthetic stories written by language models upstream (see their cards); skycmd generated none of this text.
|