---
license: apache-2.0
task_categories:
- video-text-to-text
tags:
- egocentric
- egoproactive
- video
size_categories:
- n<1K
---
# EgoProactive synthetic proactive annotations

Everything produced by the annotation and synthesis pipelines for the [AI Wearables Challenge 2026](https://wearable-ai-workshop.github.io/) EgoProactive

Dense timestamped proactive walkthroughs generated with the **ambient agent**
(orchestrator `deepseek/deepseek-v4-flash-0731` + vision `Qwen3.6-27B`),
using the held-out-validated **dense policy** (setup-phase coverage, repetition-collapse,
fire-at-onset) and a **-0.5s onset correction** at chunk-binning.

| set | clips | median events/clip | median interrupt-density | archive |
|---|---|---|---|---|
| egoconv | 234 | 24 | 53% | `egoconv_dense_annotations.tar.gz` |
| egolongqa | 158 | 40 | 51% | `egolongqa_dense_annotations.tar.gz` |

Source videos: `facebook/wearable-ai` (egoconv/val, egolongqa/val) — **disjoint from egoproactive val**.
egolongqa clips were first screened by an LLM procedural-content filter (`egolongqa_procedural.json`).

Converted per-8s-chunk training rows (`*_rows.jsonl`, `*_train.jsonl`) follow the egoproactive
format: `answers[j]` = `$interrupt$<utterance>` | `$silent$` aligned to `video_intervals`.

Measured effect (Qwen3.5-4B verbalizer, LoRA r32): egoconv-dense **+0.021 G-mean cross-domain**
(HoloAssist-140: 0.557 -> 0.577, interrupt-F1 +0.064) but **-0.042 in-domain** (egoproactive-140)
— i.e. it trades in-domain fit for out-of-distribution robustness.