megapapers-pretrain

This repo holds different preprocessed conditions of the same underlying paper set as separate configs — load a specific one with load_dataset("<repo_id>", "<config_name>").

independent

152,520 rows, one row per paper — no concatenation of citing/cited pairs.

These are all of the unique papers that show up in the megapapers project's (citing, cited) pairs: a Postgres DB scores every citation edge between two Semantic Scholar papers with an LLM-generated impact category (High/Medium/Low) — how much the citing paper's ideas, methods, or data depended on the cited paper. This config is the pool of unique papers across every edge where that score is High, restricted to the ones that also have full text available locally in peS2o v3 (peS2o only has full text for a subset of Semantic Scholar papers, and each High-impact edge needs both the citing and the cited paper to have it). Citing and cited papers are pooled together with no distinction kept, and a paper appearing on both sides of different edges is included once.

Columns:

• text (string): the paper's full text.
• id (int64): the paper's document id.

independent-shuffled

The exact same 152,520 rows as independent above — same papers, same per-paper text, same columns — just written in a randomly shuffled row order (seed 42) instead of independent's citing-then-cited, group-by-group order (each group's citing paper followed by its cited papers, in the order groups appear in fulltext_groups.jsonl). Use this to isolate whether independent's row order (which clusters a citing paper next to the cited papers it depends on) has any effect on pretraining, holding every other property of the dataset fixed.

Columns (same as independent):

• text (string): the paper's full text.
• id (int64): the paper's document id.

concat-high-impact-citing-first / concat-high-impact-citing-last

One row per citing paper: the citing paper's full text plus the full text of each of its High-impact cited papers (same High-impact edges as independent, restricted the same way to pairs with local full text), concatenated into a single document. The two configs differ only in concatenation order:

• concat-high-impact-citing-first: citing paper, then cited papers.
• concat-high-impact-citing-last: cited papers, then citing paper.

Papers are joined with a single literal </s> (Llama-2 EOS token) between them and nothing else — standard EOS-delimited pretraining concatenation, no marker text. Unlike independent, citing/cited papers are not pooled or deduplicated across rows — a cited paper referenced by multiple citing papers appears once per citing paper that concatenates it in. A citing paper with no retrievable cited-paper text is still included as a citing-only row (just its own text, no </s>).

Columns:

• text (string): the concatenated document.
• id (int64): the citing paper's document id.
• cited_ids (list of int64): document ids of the cited papers concatenated into text, in the same order they appear there.

concat-high-impact-citing-first-shuffled-seed{42,7,123} / concat-high-impact-citing-last-shuffled-seed{42,7,123}

The exact same rows as concat-high-impact-citing-first / concat-high-impact-citing-last above — same citing-paper groups, same per-row concatenated text (citing-first or citing-last order is unchanged within each row), same columns — just written in a randomly shuffled row order instead of the citing-paper order from fulltext_groups.jsonl. Analogous to independent-shuffled, but at the row (citing-paper-group) level rather than the individual-paper level: isolates whether the corpus-level row order of these two configs has any effect on pretraining, holding every other property fixed. Three seeds (42, 7, 123) are published per variant so results aren't confounded with one arbitrary permutation.

Columns (same as the corresponding non-shuffled config):

• text (string): the concatenated document.
• id (int64): the citing paper's document id.
• cited_ids (list of int64): document ids of the cited papers concatenated into text, in the same order they appear there.

concat-random-papers

Groups of 3 randomly chosen papers from the same independent pool, concatenated together with no citing/cited relationship — every paper is treated as an interchangeable, independent unit, unlike the concat-high-impact-citing-* configs above. The pool is shuffled once (seed 42) and sliced into consecutive, non-overlapping triples, so every paper appears in at most one row and no paper is reused across rows. Papers are joined with a single literal </s> (Llama-2 EOS token) between them and nothing else, in shuffle order (there is no citing/cited role to order by here).

Columns:

• text (string): the 3 papers' full text, joined with </s>.
• ids (list of int64): the 3 papers' document ids, in the same order they appear in text.

eval-random-independent

Held-out evaluation set, not a training config — 1,000 papers sampled uniformly at random (seed 42) from the full local peS2o full-text index, excluded from every training config above (disjoint from every citing/cited paper the training configs are built from). One row per paper, no concatenation. Use this to evaluate a model trained on any of the configs above, since none of its papers were seen in any of them.

Columns:

• text (string): the paper's full text.
• id (int64): the paper's document id.

eval-high-impact-suffix / eval-high-impact-prefix

Held-out evaluation sets, not training configs — the same recipe as concat-high-impact-citing-first / concat-high-impact-citing-last above (one row per citing paper, concatenated with its High-impact cited papers, joined with </s>), but built from a disjoint, later-added set of High-impact (citing, cited) edges instead: after the training corpora above were built, more High-impact annotations were added to the underlying DB, and only the papers first appearing in those new edges (never appearing anywhere — citing or cited side — in the edges the training configs were built from) were kept, then run through the same construction recipe. Each shares no papers with any training config above. The two configs differ only in concatenation order, and are named by which side of the citing paper the cited papers land on:

• eval-high-impact-suffix: citing paper first, cited papers appended after (as a suffix) — matches concat-high-impact-citing-first's order.
• eval-high-impact-prefix: cited papers first (as a prefix), citing paper appended last — matches concat-high-impact-citing-last's order.

Columns (same as their non-eval- counterparts): text (string), id (int64, the citing paper's id), cited_ids (list of int64).

eval-no-prefix-no-suffix

Held-out evaluation set, not a training config — the same disjoint, citing-paper pool as eval-high-impact-suffix/eval-high-impact-prefix above, but with no concatenation at all: each citing paper is its own independent row, with no cited papers included — neither a prefix nor a suffix of cited-paper text is appended (contrast with the two configs above, which are literally the same citing papers plus a prefix/suffix of cited-paper text). Use this to isolate how much of eval-high-impact-suffix/-prefix's eval signal comes from the citing paper alone vs. the concatenated cited papers.

Columns:

• text (string): the citing paper's full text.
• id (int64): the citing paper's document id.

Usage

from datasets import load_dataset

ds = load_dataset("hannahglz25/megapapers-pretrain", "independent")
ds = load_dataset("hannahglz25/megapapers-pretrain", "independent-shuffled")
ds = load_dataset("hannahglz25/megapapers-pretrain", "concat-high-impact-citing-first")
ds = load_dataset("hannahglz25/megapapers-pretrain", "concat-high-impact-citing-last")
ds = load_dataset("hannahglz25/megapapers-pretrain", "concat-random-papers")
ds = load_dataset("hannahglz25/megapapers-pretrain", "eval-random-independent")
ds = load_dataset("hannahglz25/megapapers-pretrain", "eval-high-impact-suffix")
ds = load_dataset("hannahglz25/megapapers-pretrain", "eval-high-impact-prefix")
ds = load_dataset("hannahglz25/megapapers-pretrain", "eval-no-prefix-no-suffix")