The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

FineWeb-Edu → Deseret Alphabet

A translation of the FineWeb-Edu sample-10BT subset into the Deseret Alphabet — a 19th-century phonetic writing system for English, encoded in Unicode at U+10400–U+1044F.

This dataset is intended for pre-training language models that operate exclusively in the Deseret Alphabet.

Dataset summary

Documents: 9,126,624 (after prose-heavy filtering)
Approximate Deseret tokens (via 8k BPE): 11.1B
Formats provided:
- fineweb-edu-deseret.jsonl — raw JSONL (125 GB), one document per line with text translated to Deseret Unicode plus original FineWeb-Edu metadata
- tokens/train.bin and tokens/val.bin — pre-tokenized uint16 token shards (21 GB + 191 MB) using the companion chrisjpatty/deseret-8k-bpe tokenizer (11.13B + 100M tokens). Drop-in for nanoGPT-style pre-training.
Source license: Open Data Commons Attribution 1.0 (ODC-By 1.0)
This derivative: also ODC-By 1.0

Each record contains:

{
  "text": "𐐜𐐲 𐐆𐑌𐐼𐐮𐐹𐐯𐑌𐐼𐐲𐑌𐐻 𐐖𐐩𐑌 ...",
  "url": "<original url from FineWeb-Edu>",
  "score": <FineWeb-Edu quality score>,
  "id": "<original FineWeb-Edu id>"
}

Translation methodology

English text was translated to Deseret with a custom pipeline:

English word → CMU Pronouncing Dictionary (g2p_en fallback for OOV)
            → ARPAbet phoneme sequence
            → Illinois Deseret Consortium SAMPA convention
            → Deseret Alphabet codepoints

Capitalization, punctuation, whitespace, and numerals pass through unchanged.

Validation

The SAMPA → Deseret stage of the pipeline was validated against the Illinois Deseret Consortium's parallel transcription of the 1869 Book of Mormon (the authoritative published Deseret text), achieving:

100.00 % byte-exact parity on the IDC spelling dictionary (1,359 entries)
99.965 % word-level parity on the full 1869 Book of Mormon (108k+ words)

The 38 residual mismatches at corpus scale are all source-data quirks in the IDC parallel transcription (SAMPA-vs-Deseret column disagreements on the same row), not algorithm bugs.

The English → SAMPA stage uses modern CMUdict pronunciations, which differ from the 1869 New England dialect in three well-understood ways:

Single-letter shorthand: 1869 Deseret used single Deseret letters as common words (Thee = "the", Bee = "be", Long-A = "a"). This pipeline writes them out phonetically.
Schwa preservation: CMUdict marks unstressed schwa explicitly; 1869 Deseret often elided them.
Cot–caught merger: Modern General American /ɑ/ vs 1869 /ɒ/ in certain words.

The translator is internally consistent and faithful to the Deseret Alphabet's documented design — it just uses modern phonology rather than 1860s New England pronunciation.

Content filtering

During fetch from FineWeb-Edu, documents were dropped if:

Fewer than 100 words
More than 2 % of characters were inside URLs
Contained common code/markup markers (```, <html>, def, function, etc.)
Fewer than 80 % alphabetic-or-punctuation characters

This kept ~94.2 % of source documents and discarded code/markup-heavy content that translates poorly to a phonetic alphabet.

Intended use

Pre-training small language models in the Deseret Alphabet
Research on phonetic-only orthographies as a modeling target
Studies of script-specific tokenization

The companion chat-format dataset for SFT is at chrisjpatty/ultrachat-deseret.

The tokenizer used for the BPE statistics quoted above is at chrisjpatty/deseret-8k-bpe.

Limitations

Uses modern American English pronunciation, not historical Deseret-era pronunciation
Quality is bounded by CMUdict + g2p_en correctness on out-of-vocabulary words
Has not been used to train production models yet (see project home for in-progress LLM)

Citation

If you use this dataset, please attribute the source FineWeb-Edu corpus:

@misc{lozhkov2024fineweb-edu,
  author = {Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas},
  title = {FineWeb-Edu},
  year = {2024},
  publisher = {Hugging Face},
}

And ideally also note the translation provenance:

FineWeb-Edu translated to the Deseret Alphabet by Christopher Patty (chrisjpatty), 2026.

Downloads last month: 31

Total file size:

156 GB