NEST

Naturalistic Early-childhood Social inTeractions — the first compositional benchmark for social reasoning in child-centric video. NEST comprises 1,052 interaction segments (~12 s each) paired with 12,197 structured questions, curated to maximize diversity across participants, developmental stages, and cultural contexts.

Each segment is manually annotated using a unified, temporally grounded schema spanning four complementary axes:

• Contextual attributes — setting, activity, age (age, location, main_activity)
• Behavioral cues — atomic actions (shared_attention, object_of_attention, event_at_time)
• Interpersonal dynamics — action-reaction contingencies (cause, consequence, reaction)
• Open narrative descriptions — free-text accounts of the interaction

This dense schema resolves the diagnostic ambiguity inherent in aggregated metrics, decomposing model performance into distinct capabilities while enabling stratified evaluation across metadata. Extensive analysis of state-of-the-art video Vision-Language Models reveals a stark performance divergence: while models demonstrate strong contextual recognition, they exhibit substantial gaps in the higher-order reasoning required to decode behavioral and interpersonal cues.

Note: We redistribute only text annotations and metadata. Source videos must be obtained directly from the original corpora; this dataset provides identifiers and frame indices to cut segments locally (see Obtaining the videos).

Privacy — FIND Dataset. A subset of NEST's annotations was collected from a clinical parent–child intervention corpus derived from the Filming Interactions to Nurture Development (FIND) program (Fisher et al., 2016), a strengths-based video-coaching intervention for caregivers of young children. For privacy reasons, the FIND Dataset is not released. The lora_ft_experiment.val split is omitted entirely, as it consists exclusively of FIND Dataset segments, leaving no public validation split.

---

Configurations

NEST is published as two configurations:

Config	Splits	Rows	Purpose
NEST	test	12,197	The full NEST benchmark — multiple-choice and open-ended questions across four question_groups.
lora_ft_experiment	train / test_nest_no_overlap	485 / 683	LoRA fine-tuning experiment. train contains unreviewed open-narrative captions; test_nest_no_overlap is the open_interactions subset of NEST restricted to source videos that do not appear in train — strict video-level leakage filter (683 questions over 214 videos / 683 segments).

lora_ft_experiment.test_nest_no_overlap is a strict subset of NEST, restricted to (a) the open_interactions question group and (b) source videos that do not appear in lora_ft_experiment.train — use it when reporting fine-tuned model performance on open-narrative questions to avoid video-level leakage from training.

The two configs share a single unified schema (15 columns). For lora_ft_experiment.train, the four NEST-specific columns (question_group, subtype, question_type, options) are null since this split contains only free-text descriptions.

Note: The lora_ft_experiment.val split is omitted from this release because it is composed entirely of FIND Dataset (private) segments — see the Privacy note above.

Loading

from datasets import load_dataset

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

# LoRA fine-tuning experiment
ft   = load_dataset("mahuynh/NEST", "lora_ft_experiment")
ft_train, ft_test = ft["train"], ft["test_nest_no_overlap"]

---

Schema

All splits share the following 15 columns. For lora_ft_experiment.train, the last four columns are null.

Column	Type	Description
question_id	string	Unique question identifier (see format below)
segment_id	string	Segment identifier, format <video_id>_seg<NN>
video_id	string	Source video, format childes<NNNN> or childplay<NNNN>
dataset	string	childes or childplay
start_frame	int	First frame of the segment in the source video (0-indexed)
end_frame	int	Last frame of the segment (exclusive)
keyframe	int	Anchor frame the question references
descriptions	JSON string	Mapping from person color label to physical description, e.g. {"red": "child in white shirt", "green": "adult with dark hair"}
prefix	string | null	Preamble to prepend to question_text when building the prompt — always includes the Use the following descriptions to identify the people… block (when descriptions is non-empty), and additionally a Critical Rule: rubric for cause / consequence / reaction. null only when both pieces are absent.
question_text	string	The question (descriptions block + any preamble stripped — see prefix)
ground_truth	string	Reference answer (option letter for MC; free-text for open-ended)
question_group	string | null	contextual, behavioral, interpersonal, or open_interactions. null for lora_ft_experiment.train.
subtype	string | null	Granular question subtype (see table below). null for open_interactions and for lora_ft_experiment.train.
question_type	string | null	multiple_choice or open_ended. null for lora_ft_experiment.train.
options	JSON string | null	List of answer options for multiple-choice questions; null for open-ended and for lora_ft_experiment.train.

For multiple-choice questions, ground_truth is the option letter ("A", "B", ...). For open-ended questions and the lora_ft_experiment.train split, it is a free-text reference answer.

NEST question_id format

question_id encodes the group + subtype: <Letter><digit>_<NNNN> (or O_<NNNN> for open_interactions).

Letter prefix	question_group	subtype	Rows
C1	contextual	age	2,196
C2	contextual	location	1,053
C3	contextual	main_activity	1,016
B1	behavioral	shared_attention	915
B2	behavioral	object_of_attention	506
B3	behavioral	event_at_time	2,328
I1	interpersonal	consequence	1,027
I2	interpersonal	cause	1,059
I3	interpersonal	reaction	1,044
O	open_interactions	(null)	1,053

Group totals: contextual 4,265 · behavioral 3,749 · interpersonal 3,130 · open_interactions 1,053.

---

Obtaining the videos

The dataset ships only annotations and identifiers — you still need to (1) download the source videos from the original corpora and (2) cut each segment using start_frame / end_frame. We provide a companion repository that automates both steps:

NEST source-video downloader — fetches CHILDES (TalkBank) and ChildPlay (YouTube) source videos and writes per-segment audio + silent MP4s in the NEST schema.

git clone https://anonymous.4open.science/r/NEST-2026-code
cd NEST-2026-code
conda activate nest
export PATH="$HOME/.deno/bin:$PATH"             # add to ~/.bashrc too

# 1. Download source videos.
python preprocessing/download/download_videos_youtube.py --cookies cookies.txt
python preprocessing/download/download_videos_childes.py

# 2. Extract per-segment clips (writes both audio and silent versions by default).
python preprocessing/download/extract_segments_from_videos.py

Source-corpus access requires:

• CHILDES — a free TalkBank account; export TALKBANK_USER / TALKBANK_PASS before running.
• ChildPlay — a cookies.txt exported from a logged-in YouTube session (YouTube blocks datacenter IPs as bot traffic).

---

Building prompts

question_text has the description block stripped. To replicate the original training prompt:

import json

descriptions = json.loads(row["descriptions"])
desc_block = "\n".join(f"- person {color}: {desc}." for color, desc in descriptions.items())

prompt = (
    "Use the following descriptions to identify the people and answer the question:\n"
    f"{desc_block}\n"
    f"Question: {row['question_text']}"
)

---

Citation

@inproceedings{huynh2026nest,
  title     = {Before Words, Beyond Speech: Evaluating Nonverbal Social Reasoning in Early Childhood},
  author    = {Huynh, Marie Amale* and Bravo-S\'anchez, Laura* and Klein, Lauren and
               Haber, Nick and Fisher, Philip and Yeung-Levy, Serena},
  booktitle = {Arxiv},
  year      = {2026},
}

The FIND Dataset is derived from the FIND program:

@article{fisher2016promoting,
  title={Promoting healthy child development via a two-generation translational neuroscience framework: The filming interactions to nurture development video coaching program},
  author={Fisher, Philip A and Frenkel, Tahl I and Noll, Laura K and Berry, Melanie and Yockelson, Melissa},
  journal={Child Development Perspectives},
  volume={10},
  number={4},
  pages={251--256},
  year={2016},
  publisher={Oxford University Press}
}

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). You are free to share and adapt the material for non-commercial purposes, as long as you give appropriate credit and distribute derivatives under the same license.

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.