SciRIFF
The SciRIFF dataset includes 137K instruction-following demonstrations for 54 scientific literature understanding tasks. The tasks cover five essential scientific literature categories and span five domains. The dataset is described in our paper SciRIFF: A Resource to Enhance Language Model Instruction-Following over Scientific Literature.
There are three dataset configurations with different max context lengths: 4096, 8192, and 16384. All experiments in the paper are performed with the 4096 context window. You can load the dataset like:
import datasets
ds = datasets.load_dataset("allenai/SciRIFF", "4096")
Dataset details
Each instance in SciRIFF has the following fields:
input: Task input (i.e. user message).output: Task output (i.e. expected model response)._instance_id: A unique id for the instance, formatted like{task_name}:{split}:{instance_id}. For instance,qasa_abstractive_qa:test:182.metadata: Metadata on the task that this particular demonstration is an instance of. More information on the schema for task metadata can be found in the SciRIFF GitHub repo.domains: Scientific field(s) that the task covers. Options include:clinical_medicine,biomedicine,chemistry,artificial_intelligence,materials_science, andmisc.input_context: Whether the input is a paragraph, full text, etc. Options include:sentence,paragraph,multiple_paragraphs(including full paper text), andstructured(e.g. code for a LaTex table).source_type: Indicates whether the input comes from a single paper or multiple. Options includesingle_source,multiple_source.output_context: Options include:label,sentence,paragraph,multiple_paragraphs,json,jsonlines.task_family: The category to which this task belongs. Options includesummarization,ie,qa,entailment, andclassification. Some categories have sub-categories which are largely self-explanatory; see the repo for more information.
License
SciRIFF is licensed under ODC-By.
Task provenance
SciRIFF was created by repurposing existing scientific literature understanding datasets. Below we provide information on the source data for each SciRIFF task, including license information on individual datasets where available. Where possible, we leveraged the excellent BigBIO collection as a starting point, rather than reprocessing datasets from scratch. In the table below, we include the name of the BigBio subset for all tasks included in BigBio; these can be loaded like datasets.load_dataset(bigbio/{bigbio_subset}).