21,069 chat examples that teach a model to rewrite a research paper for a curious 12–14-year-old: the same paper, in the authors' own voice, keeping every result, number and uncertainty, in plain words. The source papers are 4,416 open-access (CC BY) articles from PubMed Central; the rewrites are synthetic, written by two frontier models with refined prompts.
It is the training data of maximerivest/qwen3.5-9b-paper-rewriter and maximerivest/qwen3.5-4b-paper-rewriter.
Each paper gives up to five examples, one per call:
opening: from the whole paper, rewrite the title, abstract, first introduction
paragraph and conclusion.introduction_rest, methods, results, discussion: rewrite one section, given
the rewritten opening.| Column | |
|---|---|
messages | system (the task), user (the paper or section, its reference glossary, the writer label), assistant (the rewrite, in tags) |
call | opening or section |
section | which part is rewritten |
writer | writer_a (13,969 examples) or writer_b (7,100): which frontier model wrote the rewrite. The same label appears in the prompt as <writer>. |
paper_id, paper_title, paper_doi, paper_license | the source paper |
Each user message carries a reference glossary: definitions of the passage's technical
terms from an offline copy of Wikipedia and Wiktionary, as - term (long form): definition [source].
Splits: train (20,874 examples) and validation (195 examples from 40 papers). The
models' 3 benchmark papers are not in the dataset. When training, we skipped the 480
examples longer than 24,576 tokens.
The papers config lists every source paper: title, authors, journal, year, DOI,
licence and link.
papers config) and say the text was adapted.21,069 chat examples that teach a model to rewrite a research paper for a curious 12–14-year-old: the same paper, in the authors' own voice, keeping every result, number and uncertainty, in plain words. The source papers are 4,416 open-access (CC BY) articles from PubMed Central; the rewrites are synthetic, written by two frontier models with refined prompts.
It is the training data of maximerivest/qwen3.5-9b-paper-rewriter and maximerivest/qwen3.5-4b-paper-rewriter.
Each paper gives up to five examples, one per call:
opening: from the whole paper, rewrite the title, abstract, first introduction
paragraph and conclusion.introduction_rest, methods, results, discussion: rewrite one section, given
the rewritten opening.| Column | |
|---|---|
messages | system (the task), user (the paper or section, its reference glossary, the writer label), assistant (the rewrite, in tags) |
call | opening or section |
section | which part is rewritten |
writer | writer_a (13,969 examples) or writer_b (7,100): which frontier model wrote the rewrite. The same label appears in the prompt as <writer>. |
paper_id, paper_title, paper_doi, paper_license | the source paper |
Each user message carries a reference glossary: definitions of the passage's technical
terms from an offline copy of Wikipedia and Wiktionary, as - term (long form): definition [source].
Splits: train (20,874 examples) and validation (195 examples from 40 papers). The
models' 3 benchmark papers are not in the dataset. When training, we skipped the 480
examples longer than 24,576 tokens.
The papers config lists every source paper: title, authors, journal, year, DOI,
licence and link.
papers config) and say the text was adapted.