One Sentence Turns Papers into a Dataset: Paperclip Has Already Built a 20,000-Record Half-Life Dataset

James Zou of GXL has released Paperclip Datasets. What it does is straightforward: you describe the dataset you want, and it finds evidence across millions of papers, extracts fields, and organizes them into a table, producing results in minutes. The entry point is https://paperclip.gxl.ai/datasets , and the blog post is at https://gxl.ai/blog/introducing-paperclip-datasets/ .

Released alongside it is an example dataset: the All-Route Terminal Half-Life Dataset. The Hugging Face page shows that it contains 19,952 terminal half-life records from 3,270 PubMed Central articles, covering four administration-route categories: oral, iv, subcutaneous, and other. James Zou's tweet says the data was compiled from more than 2,400 drugs. The table has 14 columns, including fields such as pmcid, drug_name, normalized_drug_name, molecule_class, smiles, inchikey, administration_route, terminal_half_life, reported_half_life, summary_statistic, drug_information, cohort_information, and notes. Dataset URL: https://huggingface.co/datasets/gxl-ai/paperclip-half-life .

Beyond the numbers, what is more worth examining is how it handles trustworthiness. Each record carries a pmcid and can be traced back to the source article; reported_half_life preserves the values, units, statistics, and dispersion information from the original text; notes record unit conversions, terminal-phase determinations, and source limitations. The HF card also mentions that during the duplicate-review stage, an LLM helped compare 218 candidate pairs across 213 high-confidence groups, removing 215 redundant rows while retaining comparisons for different doses or PK occasions. Here the LLM is a cleaning aid, and the source evidence is still written into each row.

The Paperclip page also makes the usage clear. The example in the input box is Build a dataset of weight loss with GLP-1 drugs like semaglutide and tirzepatide, by drug and dose; nearby, you can adjust ROWS and AUTO-APPROVE COLUMNS, and click BUILD to start building. Preset tasks include CAR-T response rates in relapsed multiple myeloma, by product, and psilocybin trials for depression, with effect sizes. For readers without a development background, it can be understood as turning papers into a table with one sentence; for people working in AI, the focus is on the data layer: how fields are defined, how evidence can be traced back, and whether the cleaning process is auditable.

Its boundaries are also clearly stated. The HF card explicitly says that this dataset cannot be used for clinical guidance, dose selection, therapeutic monitoring, regulatory decisions, or patient-level risk assessment; half-lives under different matrices, populations, regimens, and assay methods cannot be compared automatically; and chemical identifiers are not an authoritative chemical registry. In the data, 317 records have summary_statistic set to unspecified, 3,078 lack a normalized drug name, and 7,518 lack SMILES and InChIKey. These numbers directly affect whether the table can be used for analysis, and they also affect your degree of trust in automated extraction.

In the comments, Layth says really cool awesome work, Zina Sarif says crazy cool, Cellgeist says love this, and Connor mentions cool to see the paperclip buildout continue. The feedback is very positive. The tool's real value still depends on whether the generated datasets can be reused and audited, and whether they can save those weeks of digging through papers and copying tables in specific research projects.