megapapers

megapapers is a pretraining corpus built from citing papers grouped with the full text of the high-impact papers they cite. Every group has at least one cited paper, and every paper in a group (citing or cited) has local full text in peS2o v3.

It is a separate release from hannahglz25/megapapers-pretrain, which stays unchanged. The difference is that this release has leakage-controlled train / validation / test splits: no paper appears in more than one split, whether as a citing paper or a cited paper.

Configs

The first five configs are built from the same split assignment, so each split holds the same paper texts in every one of them. Only the way the texts are arranged into rows differs. The last three keep high-impact-citing-last's citing papers, split for split, and swap in other references; see "Reference controls" below.

Config	Row =	Columns
high-impact-citing-first (default)	one group: the citing paper, then its cited papers, joined with </s>	text, id, cited_ids
high-impact-citing-last	one group: the cited papers, then the citing paper, joined with </s>	text, id, cited_ids
independent-shuffled	one paper occurrence, rows shuffled (seed 42)	text, id
concat-random	3 unrelated paper occurrences, drawn in shuffled order (seed 42) and joined with </s>	text, ids
independent-citing-only (test only)	one test citing paper, alone, in the same row order as the high-impact test splits	text, id
medium-impact-citing-last	one citing paper: some of its Medium-impact cited papers, then the citing paper, joined with </s>	text, id, cited_ids, + metadata
low-impact-citing-last	the same with Low-impact cited papers	text, id, cited_ids, + metadata
random-refs-citing-last	the same with random cited papers of that citing paper, any impact label	text, id, cited_ids, + metadata

• id / ids / cited_ids are Semantic Scholar Corpus IDs (int64). They match the peS2o id. In the concat configs id is the citing paper.
• independent-shuffled is the token-matched control. Every occurrence of a paper in a split's groups (the citing paper plus each cited paper) is its own row, with duplicates kept. A paper cited by 5 train groups appears in 5 train rows, the same number of times it appears across the concat configs.
• concat-random is the concatenation control. It uses the same occurrence pool as independent-shuffled, with papers grouped 3 per row without regard to citation links. If a split's count isn't divisible by 3, the last row is shorter, so no text is dropped.
• independent-citing-only is an evaluation set with only a test split: the 3,556 test citing papers, each once, with no cited papers. Row i is the citing paper of row i in high-impact-citing-first's and -last's test splits (same id, same text), without its cited papers.
• </s> is the Llama-2 EOS string. A few papers contain </s> literally in their own text, so don't split on it to recover individual papers. Use ids / cited_ids instead.

Rows per split

Config	train	validation	test
high-impact-citing-first	64,015	3,556	3,556
high-impact-citing-last	64,015	3,556	3,556
independent-shuffled	204,397	10,158	10,283
concat-random	68,133	3,386	3,428
independent-citing-only	—	—	3,556
medium-impact-citing-last	56,188	2,413	2,331
low-impact-citing-last	55,989	2,527	2,473
random-refs-citing-last	64,015	3,556	3,556

Per split, the total paper characters are identical across the first four configs (train 8.49B, validation 429.2M, test 421.6M). The concat configs add only the 4-character separators. The three reference controls hold different papers, so their totals differ (see below).

Reference controls

medium-impact-citing-last, low-impact-citing-last and random-refs-citing-last are counterparts of high-impact-citing-last. In each split they use the same citing papers as its split (64,015 train, 3,556 validation, 3,556 test) and the same row format, and change only which cited papers go before the citing paper:

Config	References before the citing paper
high-impact-citing-last	its High-impact cited papers
medium-impact-citing-last	its Medium-impact cited papers
low-impact-citing-last	its Low-impact cited papers
random-refs-citing-last	random cited papers of the same citing paper (any impact label, or none)

• Length matched. A row is not every reference of its kind. Its target is the word count of the same citing paper's high-impact-citing-last row in the same split (train mean about 21.4k words, about 3 papers). The usable references are shuffled (seed 42), and the row keeps the prefix, of at least one reference, whose total word count, citing paper included, is closest to the target.
• No paper crosses splits, in any config. Usable references have peS2o full text and are not the citing paper itself. A train reference appears nowhere (citing or cited) in this release's validation or test splits. A validation reference appears in no train split (of any config) and not in the test split. A test reference appears in no train or validation split (of any config). A check over all four *-citing-last configs found no paper in two splits.
• Fewer rows for Medium and Low. A citing paper with no usable reference of a kind has no row: in train 7,827 of the 64,015 for Medium and 8,026 for Low. In most of these cases the paper has Medium/Low references, but none of the cited papers has peS2o full text. Validation and test lose more (about 30% of the citing papers), because many of their Medium/Low references are already in a train split.
• random-refs-citing-last's train references are 37.4% unlabelled, 30.6% Low, 18.1% Medium and 13.9% High (validation and test: about 39% unlabelled, 22% Low, 15% Medium, 23% High), so a random row can share a reference with the high-impact row.

Config	Split	Rows	Unique papers	Refs per row (mean)	Words per row (mean / median)	Words (sum)
high-impact-citing-last	train	64,015	136,917	2.2	21,417 / 17,149	1.371B
medium-impact-citing-last	train	56,188	125,683	2.1	20,268 / 16,828	1.139B
low-impact-citing-last	train	55,989	130,754	2.2	20,972 / 17,293	1.174B
random-refs-citing-last	train	64,015	157,114	2.4	22,006 / 17,724	1.409B
high-impact-citing-last	validation	3,556	7,722	1.9	19,520 / 16,341	69.4M
medium-impact-citing-last	validation	2,413	6,212	1.7	17,533 / 15,219	42.3M
low-impact-citing-last	validation	2,527	7,094	1.9	18,645 / 16,318	47.1M
random-refs-citing-last	validation	3,556	10,190	2.1	20,042 / 16,854	71.3M
high-impact-citing-last	test	3,556	7,881	1.9	19,095 / 15,928	67.9M
medium-impact-citing-last	test	2,331	5,984	1.7	17,111 / 14,901	39.9M
low-impact-citing-last	test	2,473	6,808	1.8	18,110 / 15,771	44.8M
random-refs-citing-last	test	3,556	10,137	2.1	19,594 / 16,369	69.7M

Extra columns in these three configs:

• cited_impact (list): each reference's impact label: High, Medium, Low, None (no label) or Other (a malformed label in the source DB, 4 references in random-refs-citing-last).
• n_cited (int32): references in the row.
• n_words (int64): whitespace word count of text.
• n_available (int32): usable references of the row's kind the citing paper had.
• target_words (int64): the length target, i.e. the citing paper's high-impact-citing-last row word count.
• target_source (string): own_high_row for every row.

Source

Built from data/high_impact_pairs/fulltext_groups.jsonl (written by src/get_fulltext.py --all):

	Count
Citing papers (groups)	71,127
Unique papers (citing or cited)	152,520
Citing → cited edges	153,711

Splits

Split	Groups	Share	Unique cited	Unique papers	Edges
train	64,015	90%	81,484	136,917	140,382
validation	3,556	5%	4,770	7,722	6,602
test	3,556	5%	4,858	7,881	6,727

Why the split is not a plain random split over citing papers

Groups share papers. A cited paper can have more than 1,000 citing papers, and about 9.7k papers appear as both citing and cited. With a plain random 90/5/5 split over citing papers, about two thirds of validation and test groups would contain a paper that is also in train. Held-out loss would then measure memorization rather than generalization.

How the split is made

1. Build an undirected graph whose nodes are papers (citing or cited) and whose edges are citations. Papers whose full text is byte-identical under two different ids (12 sets) are also joined, so a duplicate can't cross splits. Self-citations (a group listing its own citing paper as cited) are dropped from cited.
2. Take its connected components. 25,581 components contain a group. The largest contains 39,174 groups (55% of the data).
3. Assign whole components to splits, so a paper can never land in two splits:
   • The largest component always goes to train. It is too large to split without cutting edges.
   • The rest are shuffled with seed 42 and assigned stratified by field of study (the citing paper's first s2fieldsofstudy entry). Validation and test each get a 5% quota per field. Components go to a held-out split only if they fit within its field quotas. Otherwise they go to train.

The result: no citing or cited paper in validation or test appears anywhere in train, validation and test share no papers, and all three splits have the same field mix:

Field	train	validation	test
computer science	33.3%	33.3%	33.3%
physics	27.1%	27.1%	27.1%
biology	8.7%	8.7%	8.7%
medicine	8.7%	8.7%	8.7%
engineering	5.2%	5.2%	5.2%
materials science	5.0%	5.0%	5.0%
environmental science	4.2%	4.2%	4.2%
mathematics	2.1%	2.1%	2.1%

Field mix is matched, but other properties may not be. Validation and test come only from outside the largest component, so they lean toward cited papers with fewer other citing papers in the corpus.

Usage

from datasets import load_dataset

ds = load_dataset("hannahglz25/megapapers", "high-impact-citing-first")
train, val, test = ds["train"], ds["validation"], ds["test"]

control = load_dataset("hannahglz25/megapapers", "independent-shuffled", split="validation")

medium = load_dataset("hannahglz25/megapapers", "medium-impact-citing-last")  # train / validation / test

Building

1. src/build_megapapers_splits.py (scripts/run_build_megapapers_splits.sh) makes the split assignment and writes data/high_impact_pairs/data_splits/{train,validation,test}.jsonl. It verifies that no corpus id or identical text crosses splits.

2. The configs are built from those JSONL splits:

   • high-impact-citing-first / -last: src/build_concat_high_impact_dataset.py
   • independent-shuffled: src/build_independent_shuffled_splits.py
   • concat-random: src/build_concat_random_splits.py
   • independent-citing-only: src/build_independent_citing_only_test.py
   • medium-impact-citing-last / low-impact-citing-last / random-refs-citing-last: src/build_impact_category_concat.py (from the high-impact-citing-last train split, the DB's citation graph and impact labels, and peS2o full text)

   Each re-reads its output and verifies that every paper occurrence is written exactly once, with unchanged text, and that no id crosses splits.

Publishing

From the repo root, on a node with internet access:

REPO_ID=hannahglz25/megapapers RELEASE_DIR=data/huggingface_megapapers \
  bash scripts/run_publish_to_huggingface.sh