Hugging Face's logo

Datasets:
jncarlo
/
tinymix

Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
1M - 10M
Tags:
small-language-models
synthetic-data
curriculum
tinystories
babylm
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
tinymix / README.md
jncarlo's picture
jncarlo
tinymix: 1000M, 100M, 10M, 1M
9cd89ed verified
preview code
|
Raw History Blame Contribute Delete
16 kB
metadata 
language:
  - en
license: other
license_name: mixed-per-source
license_link: https://huggingface.co/datasets/jncarlo/tinymix#sources-and-licensing
pretty_name: tinymix
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
tags:
  - small-language-models
  - synthetic-data
  - curriculum
  - tinystories
  - babylm
configs:
  - config_name: 1M
    data_files:
      - split: train
        path: data/1M/train*.parquet
      - split: validation_dialogue
        path: data/1M/validation_dialogue.parquet
      - split: validation_expository
        path: data/1M/validation_expository.parquet
      - split: validation_narrative
        path: data/1M/validation_narrative.parquet
  - config_name: 10M
    data_files:
      - split: train
        path: data/10M/train*.parquet
      - split: validation_dialogue
        path: data/10M/validation_dialogue.parquet
      - split: validation_expository
        path: data/10M/validation_expository.parquet
      - split: validation_narrative
        path: data/10M/validation_narrative.parquet
  - config_name: 100M
    data_files:
      - split: train
        path: data/100M/train*.parquet
      - split: validation_books
        path: data/100M/validation_books.parquet
      - split: validation_dialogue
        path: data/100M/validation_dialogue.parquet
      - split: validation_expository
        path: data/100M/validation_expository.parquet
      - split: validation_narrative
        path: data/100M/validation_narrative.parquet
      - split: validation_news
        path: data/100M/validation_news.parquet
      - split: validation_poetry
        path: data/100M/validation_poetry.parquet
  - config_name: 1000M
    data_files:
      - split: train
        path: data/1000M/train*.parquet
      - split: validation_books
        path: data/1000M/validation_books.parquet
      - split: validation_dialogue
        path: data/1000M/validation_dialogue.parquet
      - split: validation_expository
        path: data/1000M/validation_expository.parquet
      - split: validation_narrative
        path: data/1000M/validation_narrative.parquet
      - split: validation_news
        path: data/1000M/validation_news.parquet
      - split: validation_poetry
        path: data/1000M/validation_poetry.parquet

tinymix

Small, genre-balanced English pretraining corpora in four sizes (about 1M, 10M, 100M and 1000M words), built for training very small language models. The mix combines synthetic child-level stories and dialogue, simplified real text, encyclopedic and didactic text, books, poetry and news, all filtered toward clear, self-contained documents. It is inspired by TinyStories, SimpleStories, TinyHelen and the BabyLM challenge.

What is different from using those datasets directly:

  • Several kinds of text in one mix (narrative, dialogue and chat, expository, books, poetry, news), not only stories or only conversations.
  • Four sizes built with the same recipe, with complexity limits that grow with size.
  • No non-commercial sources. TinyChat (CC-BY-NC) and the BabyLM corpus (no clear license, includes subtitle and child-speech data) were left out on purpose.
  • Provenance per document (source, genre) and validation splits per genre, to see which kind of text a model is weak at.
  • The whole pipeline is included (code/) and every filter decision is documented below.

Subsets

Subset Train words Train docs ~Tokens (BPE)
1M 1.0M 4,135 ~1.3M
10M 9.96M 42,117 ~12.9M
100M 99.5M 338,230 ~129.3M
1000M 994.92M 3,040,575 ~1293.4M

Words are whitespace-separated words in train. Each subset also has about 0.5% of the documents per genre held out as validation_<genre> splits (at least one document per genre; the 1M validation splits are therefore tiny, 4 to 8 documents each, and noisy).

from datasets import load_dataset
ds = load_dataset("jncarlo/tinymix", "10M")          # splits: train, validation_narrative, validation_dialogue, ...
print(ds["train"][0])                       # {'text': ..., 'source': 'tinystories', 'genre': 'narrative'}

Columns: text (str), source (key of the table below), genre (narrative, dialogue, expository, books, poetry, news). Documents are independent; join them with an end-of-text token when training. Dialogues are formatted one turn per line (Mia: ..., or User: / Assistant: for assistant chats).

Composition

Share of training words by genre:

Genre 1M 10M 100M 1000M
narrative 38.0% 35.0% 21.0% 15.0%
dialogue 21.9% 25.0% 21.0% 19.0%
expository 40.1% 40.0% 48.0% 56.0%
books 0% 0% 8.0% 7.0%
poetry 0% 0% 1.0% 1.0%
news 0% 0% 1.0% 2.0%

Million words per source:

Source 1M (M words) 10M (M words) 100M (M words) 1000M (M words)
TinyStories V2 (GPT-4 split) 0.189 1.494 7.959 39.793
SimpleStories 0.189 1.99 7.964 59.698
Cosmopedia v0.1, stories 0 0 4.972 49.763
TinyDialogues 0.12 1.494 6.969 19.655
SODA 0.099 0.994 8.955 83.828
UltraChat 200k 0 0 3.974 72.915
SmolTalk (everyday-conversations, smol-summarize, smol-rewrite) 0 0 0.994 12.66
TinyHelen Leaner-Pretrain (book, conversation, textbook, web, wiki) 0.15 1.499 9.946 33.921
Simple English Wikipedia (20231101.simple) 0.1 0.996 9.951 21.201
English Wikipedia, lead paragraphs 0 0 0 32.885
Cosmopedia v2 (SmolLM-Corpus) 0.1 0.996 11.944 195.975
Cosmopedia v0.1 (wikihow, khanacademy, openstax) 0.05 0.497 2.984 29.767
tiny-textbooks (textbook column) 0 0 3.98 64.294
FineWeb-Edu (sample-10BT, score >= 4) 0 0 8.96 179.075
Project Gutenberg (English books) 0 0 7.96 69.649
Gutenberg Poetry Corpus 0 0 0.995 9.95
CNN/DailyMail 3.0.0 0 0 0.995 19.897

The 10M and 100M subsets are within about 1% of their targets (train plus validation splits). In the 10M build, sources that could not fill their quota with the filters (Simple English Wikipedia, Cosmopedia v0.1) were topped up by same-genre substitute sources (tiny-textbooks, Cosmopedia v2), capped per source; that is why tiny-textbooks appears there. The 1M subset keeps the 10M proportions except that Cosmopedia stories (not in the 10M mix) were replaced by TinyStories and SimpleStories.

The 1000M subset is also within about 1% of its target. Five sources ran out under the filters before their quota and were topped up by same-genre substitutes (million words): TinyDialogues by SODA (9.3), SmolTalk by UltraChat (12.3), TinyHelen by tiny-textbooks (15.9), Simple English Wikipedia by tiny-textbooks (8.7) and English Wikipedia lead paragraphs by Cosmopedia v2 (17.0).

How it was built

  1. Normalisation: ftfy, NFKC, ASCII quotes and dashes, whitespace; URLs, e-mails, phone numbers, code, LaTeX and HTML removed. Documents with any character outside ASCII, Latin-1 and Latin Extended-A letters, common punctuation and a few math symbols are rejected whole (reason charset in the build stats).
  2. Language: fastText lid.176, English with confidence >= 0.9 (skipped for TinyDialogues, whose one-word toddler turns confuse it).
  3. Quality heuristics: Gopher-style rules (mean word length, symbol and digit ratio, repeated lines, repeated bigrams, stop words), minimum length, complete ending (drops truncated generations), a short list of LLM tics (delve, tapestry, ...) and an explicit-content word list.
  4. Anti-template: no 4-word opening may exceed 0.5% of a source's documents.
  5. Deduplication across all sources: exact hash and MinHash LSH (5-word shingles, Jaccard >= 0.8).
  6. Complexity control by size (vocabulary coverage = share of lowercase tokens inside the top-N English words of wordfreq):
10M 100M 1000M
Reference vocabulary top 8k top 15k top 30k
Min vocabulary coverage 90% 90% 90%
Max Flesch-Kincaid grade (narrative, news, books) 5 8 12
Max Flesch-Kincaid grade (expository) 9 10 12
Document length 80 to 600 words 80 to 1200 words 80 to 2000 words
Dialogue: mean words per turn <= 25 <= 40 <= 80

Assistant chats (User: / Assistant:) may average up to 4 times the per-turn cap and need at least 2 turns (named dialogues need 4), because summarisation and rewrite tasks have 2 turns and assistant answers are long.

Documents longer than the maximum are split at paragraph borders. The 1M subset is a prefix slice of the 10M build; 10M, 100M and 1000M were built separately with the same recipe, so they are not guaranteed to be subsets of each other (a later fix to the end-of-document check, which wrongly rejected Wikipedia articles that end in a heading, was only applied to sources rebuilt for the 100M run). 7. Run safety nets (used to build the larger subsets without manual tuning): a probe per source picks a starting filter strictness, strictness relaxes in up to three logged steps when yield is too low, and shortfalls go to substitute sources.

Thresholds were calibrated on real data: the first values (95% coverage in the top 5k words, Flesch-Kincaid <= 5 for everything) accepted only 3% of SimpleStories and 0.6% of Cosmopedia v2, so they were relaxed as described above.

Sources and licensing

Source License Role
TinyStories V2 (GPT-4 split) CDLA-Sharing-1.0 child-level stories
SimpleStories MIT varied simple stories
Cosmopedia v0.1, stories Apache-2.0 everyday stories
TinyDialogues MIT child-directed dialogue
SODA CC-BY-4.0 social dialogue
UltraChat 200k MIT user/assistant chat
SmolTalk (everyday-conversations, smol-summarize, smol-rewrite) Apache-2.0 chat, summaries, rewrites
TinyHelen Leaner-Pretrain (book, conversation, textbook, web, wiki) MIT simplified real text
Simple English Wikipedia (20231101.simple) CC-BY-SA-3.0 / GFDL encyclopedic, simple
English Wikipedia, lead paragraphs CC-BY-SA-3.0 / GFDL encyclopedic
Cosmopedia v2 (SmolLM-Corpus) Apache-2.0 (corpus: ODC-By) synthetic textbooks, blogs, stories
Cosmopedia v0.1 (wikihow, khanacademy, openstax) Apache-2.0 procedural and didactic
tiny-textbooks (textbook column) Apache-2.0 short synthetic lessons
FineWeb-Edu (sample-10BT, score >= 4) ODC-By-1.0 educational web
Project Gutenberg (English books) Project Gutenberg License (public domain in the US) classic books
Gutenberg Poetry Corpus CC0-1.0 poetry stanzas
CNN/DailyMail 3.0.0 Apache-2.0 (articles are (c) CNN / Daily Mail) news

Licensing and attribution. This repository is a derivative compilation of the datasets above; each document keeps the license of its source. Obligations that matter when redistributing or reusing it:

  • Attribution is required for CC-BY / ODC-By / CDLA sources (listed above).
  • The Wikipedia parts (simplewiki, enwiki_intro) are CC-BY-SA-3.0 / GFDL (share-alike). If you need a share-alike-free corpus, filter out those two sources.
  • Cosmopedia, SmolTalk, UltraChat, TinyStories, SimpleStories, TinyHelen and tiny-textbooks contain model-generated text; check the terms of the generating model (GPT-4, GPT-4o-mini, Mixtral and others) for your use case.
  • CNN/DailyMail articles and FineWeb-Edu web pages are third-party content; they are included under the dataset licenses above, which do not transfer rights from the original publishers.
  • Project Gutenberg texts had the Project Gutenberg header and footer removed; they are public domain in the United States, check your jurisdiction.

No personal data was added; e-mails, phone numbers and URLs were removed, but residual personal information in web-derived sources cannot be ruled out.

Sanity-check training runs

Small GPTs (nanoGPT-style, own byte-level BPE trained on each training set). Each row shows its own parameter count and training tokens, so model sizes differ between rows. Scored in bits per byte (lower is better, comparable across tokenizers) on the held-out validation splits of this dataset. Baselines take the first 10M / 100M words of a single source with only ftfy cleaning. They are out of distribution on the genres they do not cover, so the table shows how much genre coverage matters, not a leaderboard. One seed, no tuning. Rows of different sizes are not comparable with each other. The 1M and 10M sets have no books, news or poetry split, hence n/a.

Training data Params Train tokens narrative dialogue expository books news poetry
tinymix 1M 9.59M 1.71M 1.683 1.830 2.276 n/a n/a n/a
tinymix 10M 9.59M 17.18M 0.883 0.875 1.470 n/a n/a n/a
TinyStories only, 10M words 17.41M 12.05M 1.111 2.512 2.708 n/a n/a n/a
FineWeb-Edu only, 10M words 17.41M 15.71M 1.639 1.769 1.565 n/a n/a n/a
tinymix 100M 46.43M 136.82M 0.697 0.697 0.987 1.349 1.368 1.819
FineWeb-Edu only, 100M words 46.29M 142.51M 1.331 1.312 1.215 1.726 1.670 2.306

Limitations

  • Not done: perplexity filtering with an n-gram LM, LLM-judge coherence scoring, semantic deduplication, cluster caps and benchmark decontamination. Expect some residual near-duplicates and some low-quality documents.
  • Some tinyhelen chunks start mid-thought, because the simplified web and book documents are chunks of longer texts.
  • Synthetic sources share model-specific style and topics (names, morals, classroom settings). The anti-template cap reduces this but does not remove it.
  • The toxicity filter is a short word list, not a classifier. Chat and web sources can still contain mildly sensitive content.
  • English only, mostly simple modern language; it is a poor fit for models meant to handle technical, legal or archaic text.
  • Word counts are whitespace words; token counts depend on your tokenizer.

Reproduce

code/ contains the build scripts (mix.py quotas and tiers, sources.py adapters, build.py pipeline, build_pool.py for 1000M, assemble.py, publish.py, train.py for the sanity-check runs) and requirements.txt. Building a subset: python build.py 10M (or python build_pool.py 1000M), then python assemble.py 10M. The build streams the source datasets from the Hub.

Acknowledgements

Built on the work behind TinyStories, SimpleStories, TinyDialogues, SODA, UltraChat, SmolTalk and SmolLM-Corpus (Cosmopedia, FineWeb-Edu), TinyHelen, tiny-textbooks, Project Gutenberg, the Gutenberg Poetry Corpus, Wikipedia and CNN/DailyMail. Please cite the original datasets when you use this one.