NEST

Project page · Report an issue · Citation

Naturalistic Early-childhood Social inTeractions — the first compositional benchmark for evaluating nonverbal social reasoning in child-centric video. The public release of NEST v2 contains 1,052 interaction segments (on average ~12 s) paired with 12,183 structured questions.

NEST pairs short interaction clips with structured questions and open-ended descriptions.

Each segment is manually annotated across four complementary areas:

Question group	Coverage	Questions
Contextual	Age, setting, and activity	4,257
Behavioral	Shared attention, objects of attention, and actions	3,741
Interpersonal	Causes, consequences, and reactions	3,133
Open interactions	Free-text descriptions of interactions	1,052

This repository contains text annotations and metadata. Source videos must be obtained separately.

Quick start

from datasets import load_dataset

nest = load_dataset("mahuynh/NEST", "NEST", split="test")

Each row contains one question and its reference answer, with identifiers and frame indices linking it to a video segment. Questions are multiple-choice or open-ended.

Video access

Obtain the source videos from CHILDES (TalkBank) and ChildPlay, then extract segments using start_frame and end_frame. The companion code repository provides download and extraction scripts, including access and setup instructions.

Private FIND data are not released for privacy reasons.

Evaluation reporting

For new evaluations, report:

• Inputs: frame sampling and whether audio, subtitles, or transcripts were provided.
• Multiple-choice results: accuracy by question group and subtype, the number of questions evaluated, and any answer-normalization rules.
• Open-ended results: scores separately from multiple-choice accuracy, with the scoring method and rubric. For model-based judging, record the judge model/version, prompt, and settings.
• Reproducibility: dataset and code revisions, model version, and any unavailable videos or excluded questions.

Data format

For multiple-choice questions, ground_truth is the correct option letter (A, B, etc.). For open-ended questions, it is a free-text reference answer. The descriptions and options fields contain JSON-encoded strings when populated.

When constructing prompts, prepend the stored prefix to question_text and include options for multiple-choice questions. The prefix contains person descriptions and, for interpersonal questions, additional task instructions. Rebuilding the prompt from descriptions alone omits those instructions.

Inspect a real multiple-choice example

After loading nest, run the following to display a question, person descriptions, options, and reference answer from your chosen revision. The reference answer is for inspection and scoring; do not include it in the model input.

import json

row = next(r for r in nest if r["question_type"] == "multiple_choice")
fields = (
    "question_id", "segment_id", "question_group", "subtype",
    "descriptions", "prefix", "question_text", "options", "ground_truth",
)
example = {field: row[field] for field in fields}
for field in ("descriptions", "options"):
    example[field] = json.loads(example[field]) if example[field] else None
print(json.dumps(example, indent=2, ensure_ascii=False))

Full schema — 15 columns

Column	Description
question_id	Unique question identifier.
segment_id	Segment identifier: <video_id>_seg<NN>.
video_id	Source-video identifier: childes<NNNN> or childplay<NNNN>.
dataset	Source corpus: childes or childplay.
start_frame	First segment frame, zero-indexed and inclusive.
end_frame	End frame, exclusive.
keyframe	Anchor frame referenced by the question.
descriptions	JSON object mapping person color labels to physical descriptions.
prefix	Person descriptions and task instructions to prepend to the question; may be null.
question_text	Question text, excluding the prefix.
ground_truth	Correct option letter or free-text reference answer.
question_group	contextual, behavioral, interpersonal, or open_interactions.
subtype	Question subtype; null for open_interactions.
question_type	multiple_choice or open_ended.
options	JSON list of answer options; null for open-ended questions.

Frame fields are integers; all other fields are strings or nullable strings.

Question identifiers and subtypes

question_id encodes the group and subtype: <Letter><digit>_<NNNN>, or O_<NNNN> for open_interactions.

Prefix	question_group	subtype	Questions
C1	contextual	age	2,192
C2	contextual	location	1,051
C3	contextual	main_activity	1,014
B1	behavioral	shared_attention	914
B2	behavioral	object_of_attention	505
B3	behavioral	event_at_time	2,322
I1	interpersonal	consequence	1,043
I2	interpersonal	cause	1,048
I3	interpersonal	reaction	1,042
O	open_interactions	(null)	1,052

Optional: LoRA fine-tuning experiment

The lora_ft_experiment configuration provides open-narrative training and evaluation data.

ft = load_dataset("mahuynh/NEST", "lora_ft_experiment")
ft_train = ft["train"]
ft_test = ft["test_nest_no_overlap"]

Split	Rows	Contents
train	485	Unreviewed open-narrative captions.
test_nest_no_overlap	683	Open-interaction questions from NEST, excluding source videos present in train.

Use test_nest_no_overlap when evaluating models fine-tuned on this training split. It covers 683 segments from 214 source videos and is a subset of NEST, not an additional benchmark.

Both configurations share the same schema. In train, ground_truth contains the reference description, while question_group, subtype, question_type, and options are null.

The experiment’s validation split contains only private FIND data and is omitted from this release.

Version

This dataset contains the data for the public release data of NEST v2 which maintains the 1,052 segments, improves 1,256 questions (typos, wording, distractors, etc.). The paper's results and published model scores have been computed on NEST v1.

Citation

License

The annotations and metadata in this dataset are released under CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International).

Source videos are not redistributed and retain the licenses of their original corpora — CHILDES / TalkBank terms of use, and YouTube Terms of Service for ChildPlay segments.