How the FineDNA biology text was found
The bio_* subsets of FineDNA are web pages, papers and forum posts that contain DNA sequences inside normal text.
They were found in two steps.
1. Regex filter: does the document contain DNA? (1_regex_filter.py)
We ran this filter over FineWeb-Edu, FinePDFs, PubMed Central, bioRxiv and Biostars. A document is kept when it
- contains a run of 15 or more nucleotide letters (A, T, C, G, U, N, any case),
- that run uses at least 3 different letters, so
AAAAAAAAAAAAAAAdoes not count, - the run is not inside curly braces, which in scraped text is usually LaTeX, and
- the document mentions at least one molecular-biology keyword: dna, rna, nucleotide, genome, gene, base pair, codon, primer, pcr, amplification, sequencing, nucleic acid, oligonucleotide, plasmid, transcription, translation, mrna, ribosome, trna, rrna or uracil.
python 1_regex_filter.py --input hf://datasets/HuggingFaceFW/fineweb-edu/sample/10BT --output ./dna_regex/fineweb_edu
2. p90 split: how long is the DNA? (2_p90_split.py)
Many regex hits are short: a primer, a restriction site, a motif. Training on documents with real sequences worked
better, so we measure the longest DNA word in each document with tag_dna_sequences.py: a whitespace-delimited
token of at least 6 letters that is pure A/C/G/T (or IUPAC codes, 12+ letters, at most 20% ambiguous), with
punctuation and 5'/3' markers stripped.
- Longest DNA word at least 39 bp (the 90th percentile across the regex data):
bio_p90_*. - Everything else:
bio_rest.
python 2_p90_split.py --input ./dna_regex/fineweb_edu --p90 ./p90/fineweb_edu --rest ./rest/fineweb_edu
In the Carbon-X mix, p90 gets as much weight as all of bio_rest and is read about seven times; bio_rest about
twice.
Requirements
pip install datatrove for step 1. Step 2 needs only the Python standard library and tag_dna_sequences.py, which
sits next to it.