Hugging Face's logo

Datasets:
darpa-scify
/
nsf-scify

Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
ArXiv:
Tags:
nsf
scientific-claims
science-communication
generated-data
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
nsf-scify / README.md
fallcat's picture
fallcat
Upload full NSF-SciFy generated dataset
2903557
preview code
|
Raw History Blame Contribute Delete
3.16 kB
metadata 
license: apache-2.0
language:
  - en
task_categories:
  - text-generation
tags:
  - nsf
  - scientific-claims
  - science-communication
  - generated-data
pretty_name: NSF-SciFy Full Awards
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: full
        path: data/*.parquet

NSF-SciFy Full Awards

NSF-SciFy contains 412,155 NSF award records enriched with generated nontechnical abstracts, verifiable claims, and investigation proposals.

The source technical abstracts come from the NSF awards database snapshot nsf_awards-0325.json. The generated fields were produced sequentially with the following task-specific LoRA adapters:

These adapters were trained on the materials-science plus 20K NSF-SciFy data. This repository contains their generated annotations for the full awards snapshot; it is a dataset, not a new model trained on all 412,155 records.

Load 

from datasets import load_dataset

dataset = load_dataset("darpa-scify/nsf-scify")
full = dataset["full"]
print(full[0])

Fields

  • award_id, award_year, title, directorate, division: NSF award metadata.
  • abstract: original technical abstract.
  • non_technical_abstract: parsed generated nontechnical abstract.
  • verifiable_claims: parsed list of generated verifiable claims.
  • investigation_proposals: parsed list of generated investigation proposals.
  • *_raw: raw model response retained for auditing and possible repair.
  • generation_error, *_parse_error: error details when generation or parsing failed.
  • generation_success: true only when all three parsed outputs are present and nonempty and no error marker is present.

Generation quality

Using the strict generation_success definition:

Status Records Percentage
Successful 410,313 99.553%
Unsuccessful 1,842 0.447%
Total 412,155 100%

The unsuccessful set includes overlapping categories: 1,085 records with missing, empty, or unparsed claims; 790 with missing, empty, or unparsed investigation proposals; 42 with an empty or missing nontechnical abstract; and 12 with an explicit generation error. Raw model responses are retained where available.

Important limitations

All three enriched fields are model-generated and have not been exhaustively verified by humans. A successfully parsed record can still contain factual, extraction, relevance, repetition, or writing-quality errors. Users should validate generated content for their application and use generation_success and the error fields when filtering records.

Related resources