Datasets:
dataset_info:
features:
- name: sentence1
dtype: string
- name: offensiveness
dtype: string
- name: intent
dtype: string
- name: metadata
dtype: string
splits:
- name: test
num_bytes: 3680400
num_examples: 1783
- name: validation
num_bytes: 3542986
num_examples: 1769
- name: train
num_bytes: 21927454
num_examples: 12132
download_size: 5766456
dataset_size: 29150840
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
- split: validation
path: data/validation-*
- split: train
path: data/train-*
license: cc-by-4.0
source_datasets:
- allenai/social_bias_frames
tags:
- decision-model
sbic
SBIC reliable crowd judgments, one post per row with two nullable labels.
sentence1 is the original post. offensiveness and intent are nullable categorical labels; filter out nulls for the desired task. metadata is a JSON string with the original annotation rows and aggregation statistics.
At least three distinct workers and 75% agreement on a native category are required per view (normally unanimity with three workers). Repeat stereotype rows never add votes; conflicting ratings by the same worker abstain for that field. Uncertain/probably ratings remain their original categories, not probabilities. A missing group-target answer is not a negative label. Group-target ratings and rationales are retained only in metadata: a coarse review found a unanimous but questionable negative. whoTarget=1 means identity-group targeting; the legacy Hub card reverses this mapping. Native splits are retained; normalized duplicate posts favor test, validation, then train, even when the holdout lacks a reliable label. Posts conflicting with the existing SBIC_Disagreement task splits are excluded, not moved. Metadata retains source row IDs, hashed worker IDs, original ratings, target annotations and per-view vote counts/agreement/exclusion reasons. Annotation demographics are not model inputs. Original annotations: CC BY 4.0; posts retain source terms. Rebuild: python scripts/upload_repackaged.py sbic
Original data: allenai/social_bias_frames, original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/sbic.py.
Conversion provenance: source revisions, hashes, configuration and exclusions.