SciFact: Scientific Claim Verification Dataset
SciFact pairs 1,409 expert-written scientific claims with 5,183 research abstracts labeled SUPPORTS, REFUTES, or NOINFO plus rationale sentences. The Potato config below reproduces the annotation task.
About this dataset
SciFact is a dataset for verifying scientific claims against evidence in research abstracts. It was built by Wadden, Lin, Lo, Wang, van Zuylen, Cohan, and Hajishirzi at the Allen Institute for AI and introduced in "Fact or Fiction: Verifying Scientific Claims" at EMNLP 2020.
The dataset has 1,409 scientific claims paired with a corpus of 5,183 abstracts from biomedical and other journals. Expert annotators wrote claims from citation sentences in scientific papers and introduced negations to make some claims refutable. About 809 claims form the training set; the rest are split across development and test, with the test set balanced across the three labels.
For each claim-abstract pair an annotator assigns a label of SUPPORTS, REFUTES, or NOINFO, then marks the rationale sentences in the abstract that justify a SUPPORTS or REFUTES decision. The task combines a label decision with span-level evidence identification, and it has been used to train and evaluate retrieval-plus-verification systems and COVID-19 fact-checking tools.
The Potato config below reproduces the SciFact annotation task: it shows a claim with its candidate abstract, collects the SUPPORTS/REFUTES/NOINFO label, and lets annotators select the rationale sentences that support the decision.
- Claims
- 1,409
- Abstract corpus
- 5,183
- Labels
- SUPPORTS, REFUTES, NOINFO
- Training claims
- 809
- Evidence
- Annotated rationale sentences
- Venue
- EMNLP 2020
Configuration Fileconfig.yaml
This Potato config reproduces the annotation task. Save it as config.yaml and run potato start config.yaml to try it.
# Scientific Claim Verification (SciFact-style)
# Based on Wadden et al., EMNLP 2020
# Paper: https://aclanthology.org/2020.emnlp-main.609/
# Dataset: https://github.com/allenai/scifact
#
# This task verifies scientific claims against evidence from research abstracts.
# Annotators determine if a claim is SUPPORTED, REFUTED, or has NOT ENOUGH INFO
# based on the provided abstract, and identify which sentences provide evidence.
#
# Verification Labels:
# - SUPPORTS: The abstract provides evidence that the claim is true
# - REFUTES: The abstract provides evidence that the claim is false
# - NOT ENOUGH INFO: The abstract doesn't address the claim conclusively
#
# Annotation Guidelines:
# 1. Read the claim carefully and identify what it's asserting
# 2. Read the full abstract before making a judgment
# 3. Look for sentences that directly address the claim
# 4. SUPPORTS: Evidence confirms the claim (same finding)
# 5. REFUTES: Evidence contradicts the claim (opposite finding)
# 6. NOT ENOUGH INFO: Evidence is tangentially related but doesn't verify/refute
# 7. Mark ALL sentences that provide relevant evidence (rationales)
#
# Important Distinctions:
# - "No effect found" REFUTES claims of an effect
# - Correlation claims differ from causation claims
# - Population-specific findings may not generalize
annotation_task_name: "Scientific Claim Verification"
task_dir: "."
data_files:
- sample-data.json
item_properties:
id_key: "id"
text_key: "abstract"
output_annotation_dir: "annotation_output/"
output_annotation_format: "json"
annotation_schemes:
# Step 1: Verification label
- annotation_type: radio
name: verification
description: "Based on the abstract, is the CLAIM supported, refuted, or is there not enough information?"
labels:
- "SUPPORTS"
- "REFUTES"
- "NOT ENOUGH INFO"
tooltips:
"SUPPORTS": "The abstract provides evidence that confirms the claim is true"
"REFUTES": "The abstract provides evidence that contradicts the claim (claim is false)"
"NOT ENOUGH INFO": "The abstract doesn't contain sufficient information to verify or refute the claim"
# Step 2: Evidence/rationale identification
- annotation_type: span
name: evidence
description: "Highlight the sentence(s) that provide evidence for your verification decision"
labels:
- "Evidence"
label_colors:
"Evidence": "#22c55e"
tooltips:
"Evidence": "Sentences that directly support or refute the claim"
allow_overlapping: false
# Step 3: Confidence
- annotation_type: likert
name: confidence
description: "How confident are you in your verification?"
min_value: 1
max_value: 5
labels:
1: "Very uncertain"
2: "Somewhat uncertain"
3: "Moderately confident"
4: "Confident"
5: "Very confident"
allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 2
allow_skip: true
skip_reason_required: false
Sample Datasample-data.json
[
{
"id": "sci_001",
"claim": "Vitamin D supplementation reduces the risk of respiratory infections.",
"abstract": "Background: Vitamin D has immunomodulatory effects that may protect against respiratory infections. Methods: We conducted a meta-analysis of 25 randomized controlled trials including 11,321 participants. Results: Vitamin D supplementation reduced the risk of acute respiratory infection (OR 0.88, 95% CI 0.81-0.96). The protective effect was strongest in participants with baseline vitamin D deficiency. Conclusion: Vitamin D supplementation is safe and provides modest protection against respiratory infections."
},
{
"id": "sci_002",
"claim": "Coffee consumption increases the risk of cardiovascular disease.",
"abstract": "Objective: To examine the association between coffee consumption and cardiovascular disease risk. Methods: Prospective cohort study of 468,629 UK Biobank participants followed for 11 years. Results: Compared to non-coffee drinkers, moderate coffee consumption (3-4 cups/day) was associated with lower risk of cardiovascular disease (HR 0.85, 95% CI 0.79-0.92). Heavy consumption (>6 cups/day) showed no significant association. Conclusion: Moderate coffee consumption is associated with reduced cardiovascular risk."
}
]
// ... and 4 more itemsTry it live — no install
Boot the real Potato server in your browser (WebAssembly) and annotate with this exact config. Nothing leaves your machine.
▶ Run live in your browserGet This Design
Clone or download from the repository
Quick start:
git clone https://github.com/davidjurgens/potato-showcase.git cd potato-showcase/text/fact-verification/scientific-claim-verification potato start config.yaml
Dataset & paper
Wadden et al., EMNLP 2020
Citation (BibTeX)
@inproceedings{wadden-etal-2020-fact,
title = "Fact or Fiction: Verifying Scientific Claims",
author = "Wadden, David and Lin, Shanchuan and Lo, Kyle and Wang, Lucy Lu and van Zuylen, Madeleine and Cohan, Arman and Hajishirzi, Hannaneh",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2020.emnlp-main.609",
pages = "7534--7550"
}Details
Annotation Types
Domain
Use Cases
Tags
Found an issue or want to improve this design?
Open an IssueRelated Designs
ESA: Error Span Annotation for Machine Translation
Error span annotation for machine translation output. Annotators identify error spans in translations, classify error types (accuracy, fluency, terminology, style), and rate severity.
LongEval: Faithfulness Evaluation for Long-Form Summarization
LongEval is the EACL 2023 protocol for human evaluation of faithfulness in long-form summaries (Krishna et al.). This Potato config reproduces its fine-grained, clause-level faithfulness judgments against source documents.
News Headline Emotion Roles (GoodNewsEveryone)
Annotate emotions in news headlines with semantic roles. Based on Bostan et al., LREC 2020. Identify emotion, experiencer, cause, target, and textual cue.