Datasets:
dataset_info:
features:
- name: id
dtype: string
- name: title
dtype: string
- name: context
dtype: string
- name: question
dtype: string
- name: answers
struct:
- name: text
list: string
- name: answer_start
list: int32
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 18506410
num_examples: 9874
- name: validation
num_bytes: 2091143
num_examples: 1182
- name: test
num_bytes: 2022307
num_examples: 1201
download_size: 2625032
dataset_size: 22619860
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
license: apache-2.0
source_datasets: []
tags:
- decision-model
friendsqa
FriendsQA span QA over Friends transcript scenes, native trn/dev/tst partitions.
The official release gives answers as utterance_id + an inner token span (or, for
"who is speaking" questions, the speaker name with no span). Each dialogue is flattened to one
Speaker, Speaker: utterance line per utterance, and every answer is relocated to a character
offset in that flattened context, verified verbatim (simultaneous multi-speaker lines locate
correctly because the gold speaker name is always a prefix of the joined speaker list). Only 2
of 16,662 train answers carry a genuine source token-offset bug and cannot be relocated; those
two gold spans are dropped and every other valid gold answer for their question is kept, so no
question is dropped. Duplicate (text, offset) answers from different annotators are collapsed to
one gold span. Original terms: Apache License 2.0 (repository LICENSE.txt).
Original data: original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/dialogue.py.
Conversion provenance: source revisions, hashes, configuration and exclusions.