Datasetpublic

ambient-intelligence-labs/egolongqa-synth-annotations

EgoLongQA synthetic MCQs, teacher traces and annotation outputs Everything produced by the annotation and synthesis pipelines for the AI Wearables Challenge 2026 EgoLongQA ≤2B track, other than the distillation set (which lives in infinitylogesh/egolongqa-junior-distill). ⚠️ Read this before counting rows The synthetic set is 943 questions over 408 videos, and it is stored two ways: file rows shape training_sets/train_synth_v3.jsonl 943 flat — one row… See the full description on the dataset page: https://huggingface.co/datasets/ambient-intelligence-labs/egolongqa-synth-annotations.

sourceHugging Faceotherupdated 17d agoView on Hugging Face

0likes127downloads

Dataset Card

EgoLongQA synthetic MCQs, teacher traces and annotation outputs

Everything produced by the annotation and synthesis pipelines for the AI Wearables Challenge 2026 EgoLongQA ≤2B track, other than the distillation set (which lives in infinitylogesh/egolongqa-junior-distill).

⚠️ Read this before counting rows

The synthetic set is 943 questions over 408 videos, and it is stored two ways:

file	rows	shape
`training_sets/train_synth_v3.jsonl`	943	flat — one row per question (the training form)
`annotations/egolongqa_synth_v3.jsonl`	408	nested — one row per video, each with a `questions` list

Both contain the same 943 questions. synth_questions is the default config and shows the flat 943. A raw row count on the nested file gives 408, which is videos, not questions.

Configs

config	contents
`synth_questions` (default)	the 943 flat synthetic training rows
`synth_by_video`	the same data nested per video, plus v2, and the activity / signage styles
`agent_annotations`	Ambient-agent outputs over egoconv, plus pilots
`teacher_traces`	oracle-filterable teacher passes (luna-pro, qwen-max hard slice)
`teacher_traces_122b`	qwen122b hard slice (different schema, hence its own config)
`rejects`	rejected generations with raw text — see below
`distill_train`	the 605-row real-trace training set, plain and option-shuffled
`eval_egolongqa`	`eval70` and its debiased twin `eval70_shuf`
`eval_ood`	500 human-written QAEgo4D-MC questions, near-uniform gold

Why this data is worth keeping despite a negative headline

In-distribution the synthetic set added nothing — five controlled arms, all null or worse.

Out of distribution the ranking reverses and it is the best of three:

model	QAEgo4D acc	on parsed	predicted C-rate (gold 27%)
base 2B	48.2%	54.6%	17%
distill on real traces	54.6%	55.4%	38% — imports the val skew
synth-only, 943 balanced Q	56.6%	58.4%	25% — calibrated

Synthetic data teaches the task (27.1% → 54.3% from synthetic alone, on unseen videos, with labels no human wrote) and produces a calibrated model. It simply cannot show that on a benchmark whose metric rewards the prior it removes.

Caveat on the negative arms: real rows carry a median of 5 distinct teacher descriptions; every synthetic row carries 1. Cutting descriptions from ~5 to ~1.6 cost 17.2pp in a separate controlled test. The arms never controlled for this, so the transfer question is open.

Provenance

Generated by the Ambient agent (https://github.com/ambient-intelligence-hq/ambient) over Ego4D video: VLM description → LLM question generation → adversarial rejection → frame-grounded verification → letter balancing. Raw generator output was 42% "A" before balancing.

rejects/ is kept deliberately: rejects had originally been counted but not kept, which hid two parser bugs. Logging raw text exposed both, and re-parsing salvaged 141 traces from already-paid-for API calls.

No human has verified any generated label. An independent model agreed with gold on 41/41 questions where it produced an answer — evidence, not proof.