cmeister/tokenizer-lm-corpora

tokenizer-lm training corpora

The two corpora behind cmeister/tokenizer-lm-ablations: the mixture every released language model trained on, and the pool the tokenizers were trained from. Both are samples of public corpora, released so the models' results can be reproduced rather than taken on faith.

Staged 2026-10-19.

lm_training_mixture/

100 parquet shards, 5,000,000 documents, 24.97 GB of text, 13.71 GB on disk. Sampled with seed 42; metadata.json records every source path, weight, document count and byte count, and lm_dataset_sampler.py is the script that produced it.

Composition by sampling weight: FineWeb-Edu 36.8%, FineWeb-2 31.6% across 30 languages, FineMath 15.8%, StarCoder python and javascript 15.8%.

Read this before using it

The shards have one column, text. No document id, no URL, no source dataset, no repository name. Every provenance field was dropped when the mixture was sampled, and the sampling was not recorded per document.

The practical consequences are worth stating rather than leaving to be discovered:

• The 15.8% of the mixture that came from StarCoder cannot be identified after the fact, so it cannot be separated from the prose, attributed to the repositories it came from, or removed on request. If you maintain a repository that was in The Stack and you have opted out, we cannot honour that opt-out inside this file, because the information needed to act on it no longer exists. We can remove the corpus as a whole; please open a discussion on this repository.
• Per-language analysis of the multilingual portion is not possible from the file alone.
• Deduplication against another corpus has to be done on text, not on identifiers.

The tokenizer pool below does not have this problem, and is the better starting point for anything that depends on knowing where a document came from.

tokenizer_training_pool/

1,921 parquet files over 6 sources, 3.21 GB. This is the pool the tokenizer-training compositions draw from; a given tokenizer uses a weighted subset of it, with the weights recorded in scripts/train_custom_tokenizers.py in the study's code repository. sampling_stats.json records the per-source targets and sampling temperatures, and tokenizer_dataset_sampler.py is the script that produced it.

Source	Files	GB	Columns
finemath	1	0.05	id, metadata, text
fineweb	50	1.41	id, metadata, text
fineweb2	1,781	1.36	id, metadata, text
infimath	1	0.05	id, metadata, text
megamath	1	0.11	id, metadata, text
starcoder	87	0.22	clarity, content, difficulty, educational, id, max_stars_count, max_stars_repo_name, max_stars_repo_path, output, practice, quality, quantile

Provenance is intact here. StarCoder rows carry max_stars_repo_name and max_stars_repo_path, so code can be attributed to its repository and removed on request. Web and math rows carry an id and a metadata struct holding the Common Crawl path and the source domain.

Where this came from, and what that means for licensing

The underlying public datasets are FineWeb-Edu (ODC-By), FineWeb-2 (ODC-By), FineMath (ODC-By) and StarCoderData, whose card lists its license as "other" and which carries The Stack's opt-out undertaking. Licenses as stated on each dataset's card, read 2026-08-17.

What is published here is not sampled from those public versions directly. It was sampled from quality-filtered derivatives prepared inside the Swiss AI initiative, so this release also redistributes that filtering work. Attribution for the filtering belongs to that project.

If you are an author or maintainer with a removal request, open a discussion here. For the tokenizer pool we can act on it precisely; for the language-model mixture, see the limitation stated above.

Citation

See the model repository cmeister/tokenizer-lm-ablations for the study these corpora belong to.