petadex/petai-labels-llm

PETai Labels, LLM Extracted

4,694 activity labels on 455 enzyme sequences, extracted by an LLM pipeline from 5,960 extraction files covering the plastic-degradation literature. 572 papers yielded at least one usable value. Every row carries the sentence from the paper the value was taken from. This is the llm source of petadex/petai-activity-labels.

These labels are not validated against a human-read gold set. Treat them as weak supervision, and read the warnings below before using them for anything else.

Why it exists

The human-curated set covers 59 papers thoroughly. This covers 572 papers shallowly, and it is where most of the release's sequence breadth comes from: 358 of its 455 sequences carry no label from any other source. The price is that nobody read the paper.

Files

File	Rows	What it is
labels.csv	4,694	Long format, the shared schema the unified release merges
measurements.csv	4,694	The same rows with the extraction trail: source file, protein id, assay id, raw value, where in the paper it came from, and the evidence sentence
sequences.fasta	455	The sequences, named by MD5
manifest.json		Counts and input checksums

source_row_id is <source file stem>|<index in the file's activity list>, so every unified label resolves to one row here and to one entry of one extraction JSON.

measurements.csv columns beyond the shared schema

file, json_doi, doi_from, protein_id, protein_name, assay_id, value_raw, data_source_tags, evidence_sentence.

evidence_sentence is populated on all 4,694 rows, median 172 characters. It is the one thing that makes these labels checkable: read the sentence and you can tell within seconds whether the value, the unit and the protein were lifted correctly.

Read before training

• There is no controlled vocabulary. 4,694 rows carry 2,028 distinct variable strings and 585 distinct unit strings, because the model wrote whatever the paper said. The most common units are percent 865, kcal/mol 353, mM 259, % 206, and % and percent are the same thing. Any use starts with mapping these onto a target of your own.
• No assay conditions. ph, temperature_c and time_h are empty on every row. The extraction did not capture them. If a model needs conditions, the human-curated and kinetics sources are the ones that have them.
• No negatives. Zero of 4,694. A pipeline prompted for activity values reports activity, so the absence is evidence of extraction bias rather than of biology. The release's negatives come from petadex/petai-halo-screen and the human-curated set.
• Substrate is missing or unresolved on 91 rows, where it reads NOT_FOUND.
• The DOI is not always the paper's. The pipeline sometimes filled doi with a cited paper's DOI or a supplement's, so the file name is trusted over the JSON field where they disagree. doi_from records which was used: json on 4,442 rows, filename on 252.
• Validation is possible but has not been run. 58 of the 455 sequences also appear in a human-curated or kinetics source, which is enough overlap to measure extraction error on values where a person read the same paper. That comparison is the obvious next step and no number from it exists yet.

Where in the paper each value came from

data_source_tags is multi-valued, pipe separated:

Tag	Rows
Main body	4,039
Text/captions	1,606
Table	1,592
Figure (exact)	938
Supplement	236
Figure (approximated)	97
Figure (approximate)	1

The last two are one tag spelled two ways. Values read off a figure are estimates and are worth separating from table and text values if precision matters.

Deduplication

Two passes, both of which changed the counts against the first unified release (4,898 labels, down to 4,694 here):

• Papers extracted twice under both Angewandte editions are kept once: an _ange. file is skipped when its _anie. twin is present.
• Rows repeating the same sequence, variable, unit, substrate, value and paper are kept once.

Licensing

CC BY 4.0, which covers what is ours to license: the compilation, the extraction pipeline's output structure, the sequence resolution and the derived columns. CC BY 4.0 also licenses the database right in the compilation where that right exists.

What it does not cover:

• evidence_sentence is quoted material. Each value is accompanied by one short sentence reproduced verbatim from the source paper, included so the extraction can be verified rather than trusted. Those sentences are quotations from the cited works, not our content, and the CC BY grant does not extend to them. They are short, factual and attributed to a DOI on every row, which is what quotation for verification normally requires. If you redistribute this column, you are redistributing quotations and the same conditions apply to you. Drop the column and the rest of the dataset is unaffected.
• The measured values themselves are facts reported in the cited papers. Cite the DOI in ref when you use a value.
• No source PDF is redistributed here.
• Sequences resolved against GenBank, RefSeq, UniProt and RCSB carry those services' own terms. UniProt is CC BY 4.0.

Citation

@misc{petai_labels_llm_2026,
  title  = {PETai Labels, LLM Extracted},
  author = {{iGEM Toronto 2026}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/petadex/petai-labels-llm}
}

Staged by build_source_repos.py from the enriched extraction JSONs.