spadeMIA/BioMedical_Corpus_1024_2040

PMC 1024-2040 Biomedical Fine-Tuning Corpus

Summary

This is a cleaned biomedical long-text corpus for autoregressive language-model fine-tuning, held-out evaluation, and membership-inference experiments.

split	rows	role
train	10,000	fine-tuning (membership-positive population)
test	1,000	held-out (membership-negative population)
evaluation	700	balanced membership-inference set: 350 members and 350 non-members

The test split is the full 1,000-row held-out pool in its original order. The evaluation split holds exact copies of 350 train rows (members, label = 1) and 350 rows from a separate part of the same source scan that was used for neither train nor test (non-members, label = 0). See Evaluation Split.

The public schema has two columns:

text:  large_string
label: int64   (train: all 1, test: all 0, evaluation: 1 = member, 0 = non-member)

No PMCID, date, license, URL, or provenance fields are included in the public dataset files.

Token Contract

The corpus is built for EleutherAI/pythia-2.8b.

Token counts were computed with:

tokenizer.encode(text, add_special_tokens=False)

requirement	value
stored text minimum	1024 tokens
stored text maximum	2039 tokens
EOS token budget	+1 token
maximum sequence budget	2040 tokens
EOS stored inside text	no

Final token-count statistics:

stat	train	test	evaluation members	evaluation non-members
rows	10000	1000	350	350
min	1025	1048	1036	1038
mean	1953.9	1950.8	1941.1	1961.8
median	2002	2000	2003	2005.5
max	2039	2039	2039	2039
p1	1228.0	1185.8	1172.9	1280.9
p5	1640.0	1644.0	1517.9	1666.0
p25	1966.0	1960.0	1958.0	1974.0
p75	2021.0	2021.0	2022.0	2020.8
p95	2035.0	2034.0	2036.0	2032.5
p99	2039.0	2039.0	2039.0	2037.0
share at >= 2030 tokens	12.7%	12.0%	14.6%	10.3%

Source

Source dataset:

common-pile/pubmed_filtered

Rows were filtered to:

filter	requirement
publication date	publication_date >= 2020-09-01
language	English
token length	at least 1024 Pythia tokens after cleaning
license class	CC-BY, CC-BY-SA, or CC0

The Hugging Face metadata uses license: other because the corpus contains a mixture of CC-BY, CC-BY-SA, and CC0 rows.

Documents are selected after a date-based filter intended to reduce overlap with known Pythia training sources: PubMed Central is a named component of the original 2020 Pile, and documents dated before 2020-09-01 were excluded. This is a documented, verifiable filter, not a guarantee that selected documents are absent from every language-model pretraining corpus.

Cleaning

Cleaning was deterministic and designed to remove PMC serialization artifacts without damaging biomedical prose, chemistry, units, or real math notation.

Removed or normalized artifact classes include:

artifact class	example
Pandoc anchors	{#sec1}
escaped citation groups	\[[@bib1]\]
numeric bracket citations	\[1, 2\]
bare citation markers	[@CR40]
empty citation parentheses	(), (;), (--)
empty superscript citation shells	^^, ^,^, ^--^
cross-reference links	[Figure 1](#fig1){ref-type="fig"}
LaTeX document wrappers	\documentclass, \begin{document}
raw HTML/XML tags	<p>, </italic>
Word field-code remnants	MERGEFORMAT
confirmed escaped literal punctuation	\*, \_, \#
markdown image placeholders	
generic markdown links	[label](https://example.org)
Pandoc smallcaps spans	[TEXT]{.smallcaps}
orphan bracket shells	[[ word, []
dense markdown-table pipes	pipe-separated table rows

Ambiguous doubled backslashes and escaped parentheses are deliberately not auto-stripped because they can be legitimate LaTeX, math, or prose notation.

Truncation

Each selected row is a deterministic prefix of the cleaned parent document. The pipeline uses a three-tier boundary hierarchy:

boundary type	train	test	evaluation members	evaluation non-members
paragraph	8	1	2	0
sentence	9965	997	347	350
token-boundary cutoff	27	2	1	0

Processing Summary

stage	rows
rows scanned	27,000
eligible after filtering (original build)	24,846
unique eligible after deduplication (original build)	24,843
selected train rows	10,000
selected held-out rows (published as test)	1,000
unused eligible rows (original build)	13,843
evaluation non-members drawn from the unused rows	350
evaluation members drawn from train	350

The evaluation non-members were processed by re-running the saved cleaning pipeline on the same 27,000 scanned rows. That run accepts 24,831 rows (24,828 unique, 13,840 of them not among the 11,000 rows selected for train and test), slightly different from the counts of the original build above.

Date Distribution

publication year	train	test	evaluation members	evaluation non-members
2020	1144	109	39	39
2021	2307	234	91	97
2022	2537	248	102	72
2023	3165	323	90	114
2024	847	86	28	28

Minimum selected publication date:

2020-09-01

License Distribution

license class	train	test	evaluation members	evaluation non-members
CC-BY	9962	996	349	350
CC-BY-SA	6	2	0	0
CC0	32	2	1	0

Evaluation Split

The evaluation split supports membership-inference experiments against models fine-tuned on train. Its 700 rows are 350 members and 350 non-members.

Members. Exact copies of 350 train rows, drawn with np.random.default_rng(42).permutation(10000)[:350] (indices into train).

Non-members. Exact copies of 350 articles from the same scan of common-pile/pubmed_filtered (the first 3,000 lines of shards 0008 to 0016) that were not selected for train or test. The scanned rows whose PMCID is not one of the 11,000 train and test PMCIDs were put in the order given by np.random.default_rng(20261006).permutation, and the first 350 rows that passed every check below were kept:

• the cleaning, date (>= 2020-09-01), license (CC-BY, CC-BY-SA, CC0) and English filters, and a final length of 1,024 to 2,039 EleutherAI/pythia-2.8b tokens;
• no PMCID, exact text, or whitespace- and case-normalized text shared with train, test, or the rows already selected;
• a word 5-shingle Jaccard similarity below 0.7 to every train row, every test row, and every row already selected (MinHash with 128 permutations and locality-sensitive hashing to find candidate pairs, then exact Jaccard).

No title-based check is used, because most of these texts begin with a section heading such as "1. Introduction" and not with an article title.

Processing of the non-members. The train and test rows were built in two steps: an initial build, then a repair pass that applied further cleaning rules (empty citation containers and leftover markup from the source conversion) to the already truncated text. The non-members were processed the same way: the text is first cut to the token budget with those later rules switched off, and the full cleaner is then applied to the cut text. On a seeded sample of 400 published rows this reproduces 380 rows exactly and gives the same truncation-tier counts as the published rows (1 paragraph, 397 sentence, 1 token-boundary cutoff). The mean length difference on that sample is +9.7 characters.

The 700 rows are stored in a fixed random order, np.random.default_rng(20261007).permutation(700) applied to the 350 member rows (in their order in the previous revision of this split) followed by the 350 non-member rows (in selection order), so that row position carries no label information:

row-position check	value
AUROC of row index vs. label	0.538
AUROC of reversed row index vs. label	0.462
permutation p-value (20,000 shuffles)	0.084
runs test p-value	0.940

Members are, by construction, also present in train. Non-members are not in train and not in test.

Balance and Leakage Checks

Members and non-members of the evaluation split were compared on token count, character count, pre-truncation token count, truncation tier, publication year and month, license class, source shard, and newline count (KS, Mann-Whitney, and chi-square tests; 15 tests). None is significant after Holm correction; the smallest raw p-value is 0.016 (pre-truncation token count, KS), which is 0.25 after correction. For example, token count: KS p = 0.25, Mann-Whitney p = 0.45; publication year: chi-square p = 0.085.

Because evaluation fragments are taken by percentage of each document, a mismatch in document length would shift fragment lengths. For first, middle, and last fragments at 5, 10, 25, 50, 75, and 100 percent, the AUROC of fragment length alone against the label is between 0.478 and 0.484, and every 95% interval contains 0.5.

Text-only baselines that never query a language model, evaluated on the 700 evaluation rows (parent-level bootstrap, 5,000 resamples, 95% intervals):

baseline	AUROC (member = higher value)	95% CI
character count	0.479	0.435 - 0.522
zlib compressed length	0.484	0.442 - 0.527
zlib compression ratio	0.529	0.486 - 0.572
token count	0.484	0.440 - 0.527

Five fitted baselines (five-fold out-of-fold logistic regression) were also run. The permutation p-values come from 2,000 label shuffles (1,000 for the TF-IDF baseline) with the same procedure, and Holm correction covers the five fitted baselines together with the raw token-count test:

fitted baseline	mean AUROC over 30 CV seeds (range)	permutation p	Holm p
token count	0.518 (0.509 - 0.527)	0.269	1.000
character count plus zlib ratio	0.506 (0.484 - 0.526)	0.878	1.000
all metadata features	0.526 (0.501 - 0.549)	0.106	0.639
eight markup-residue counts	0.470 (0.434 - 0.501)	0.755	1.000
TF-IDF unigrams (5,000 features, fitted inside each fold)	0.496 (0.467 - 0.523)	0.798	1.000

None of these is distinguishable from chance after correction. In the single fixed-seed literal check the all-metadata baseline has AUROC 0.550 with a bootstrap interval of [0.508, 0.592], which excludes 0.5; its permutation p-value is 0.106. The permutation tests were added to this card in an earlier revision, after a fixed-seed interval for one fitted baseline excluded 0.5. A text-only baseline should still be reported alongside any attack result.

Duplicate Checks

check	result
exact duplicate texts within any split	0
duplicates after case-folding and whitespace normalization	0
duplicate source documents (parent IDs)	0
train / test text overlap	0
evaluation non-members present in train or test	0
evaluation members present in train	350 of 350 (by design)
near-duplicates (5-word-shingle Jaccard >= 0.7 or >= 0.9) within train	0
near-duplicates between train and the held-out pool	0
near-duplicates between evaluation non-members and train, test, or each other	0 (no candidate pairs; an exact check of 40 random non-members against all 11,000 train and test rows found a highest Jaccard of 0.005)

Near-duplicate candidates were generated with MinHash and locality-sensitive hashing and verified with exact Jaccard similarity over word 5-shingles.

Verification

The local review notebook re-scans every exported row of train and the held-out pool, and the same strict detectors were run on the 350 evaluation non-members. All detectors report zero affected rows:

strict artifact class	train	test	evaluation non-members
residual pandoc anchors	0	0	0
escaped-bracket citations	0	0	0
numeric bracket citations	0	0	0
dash-only citation remnants	0	0	0
LaTeX document wrappers	0	0	0
raw HTML/XML tags	0	0	0
cross-reference links	0	0	0
bare citation markers	0	0	0
Word MERGEFORMAT artifacts	0	0	0
empty citation parentheses	0	0	0
empty superscript citation shells	0	0	0
escaped literal punctuation	0	0	0
markdown image placeholders	0	0	0
Pandoc smallcaps markup	0	0	0
orphan bracket shells	0	0	0
dense markdown-table pipes	0	0	0
broken citation keys or bibliography remnants	0	0	0
Pandoc underline wrappers	0	0	0
broken-URL markdown links	0	0	0

Token counts were recomputed with the EleutherAI/pythia-2.8b tokenizer for all train, test, and evaluation rows and agree with the stored counts.

Ordered sequence hashes over the public text column (SHA-256 over each UTF-8 text prefixed by its byte length as an 8-byte big-endian integer):

train_text_sequence_sha256      = 3f5767c782aca43da42cce8b456afb1e95e100b73c0abf4d4406e7d3b273338a
test_text_sequence_sha256       = e5207234ecbd3e80a41348a10da42a6c3ad36bba1e3c612e70f3dbbded77d37d
evaluation_text_sequence_sha256 = b46763873162f7b2b35ba4561e16724f131cdc8a78e90cd53495f0888a817bf4

Known Limitations

• Text-only residue of the source conversion remains. Inline Pandoc markup and embedded LaTeX math were retained deliberately, because removing them risks damaging real notation. The residue appears in both evaluation classes at comparable rates, with some differences (for example dollar-sign math delimiters, 12.0% of member rows and 7.7% of non-member rows). Per class (350 rows each), the share of rows that contain each pattern is:

  pattern	members	non-members
  Pandoc superscript, ^2^-style	44.9%	47.1%
  Pandoc subscript, ~2~-style	28.0%	29.4%
  Pandoc bold, **text**	23.1%	27.1%
  dollar-sign math delimiters ($)	12.0%	7.7%
  no-break space (U+00A0)	35.1%	39.7%
  thin space (U+2009)	25.1%	26.0%
  narrow no-break space (U+202F)	0.3%	0.3%
  blank line	23.1%	24.0%

  These were counted, not removed. Unicode spaces from the source conversion are also retained. Models fine-tuned on this corpus will see this markup, and it is present in train as well as in held-out rows.

• The date filter reduces, but does not rule out, overlap with language-model pretraining data.

• The evaluation non-members were processed with a procedure reconstructed from the build history (see Evaluation Split); it reproduces 380 of 400 sampled published rows exactly, not all of them.

• In the fixed-seed literal check the all-metadata baseline reaches an AUROC of about 0.55; the permutation test does not find it distinguishable from chance (see Balance and Leakage Checks).

• Rows are deterministic prefixes of longer articles and begin at the start of the document.

• With 350 non-members, strict low-false-positive-rate operating points (for example 1% FPR) are resolved by only a few examples.

Loading

from datasets import load_dataset

ds = load_dataset("spadeMIA/BioMedical_Corpus_1024_2040")
train, test, evaluation = ds["train"], ds["test"], ds["evaluation"]

Usage

Fine-tune on train only. Do not use test or evaluation for gradient updates, early stopping, checkpoint selection, or attack hyperparameter tuning. When scoring partial fragments, keep the label of the full parent record.

Revision Notes

This revision restores test to the full 1,000-row held-out pool and replaces the 350 evaluation non-members. In the previous revision the non-members were taken from the held-out pool and removed from test (650 rows); they are now 350 records from the unused part of the same source scan, and the former non-members are back in test. The 350 members, the train split, and the train row order are unchanged. The evaluation split has a new fixed random row order. The earlier revision changed only the row order of evaluation and the documentation. The repository was renamed from spadeMIA/pmc_finetune_corpus_1024-2040_tokens.

Reproducibility

field	value
preprocessing notebook	pmc_1024_2040_preprocessing.ipynb
review notebook	pmc_1024_2040_dataset_review.ipynb
artifact repair notebook	pmc_1024_2040_artifact_repair.ipynb
non-member selection script	pmc_1024_2040_nonmember_rebuild.py
preprocessing version	pmc_1024_2040
corpus seed	42
member draw seed	42
non-member selection seed	20261006
evaluation row-order seed	20261007
tokenizer	EleutherAI/pythia-2.8b
tokenizer vocabulary SHA-256	28166a9e496c12e6772d83b5b03dbeacae6d3259b138649bf6e3f54367493f6e
final audit timestamp	2026-10-07T21:28:33.187881+00:00

Use this corpus only under the token and cleaning contract documented above.

Authors

Curated by Batu Koray Masak at the Security, Privacy and Data Engineering (SPADE) Lab, Koç University.