zhizeng115/EchoScout-FourTask

EchoScout Four-Task Search

Four frozen-scene search tasks: visual object search (HOS), sound-category search, sound-duration search, and sound-onset-order search. Each episode requests one object and permits up to ten rotate/submit decisions. This is the earlier 10-second/frozen-scene benchmark, not the later 30-second Active AVQA benchmark.

Task	Train	Validation	Test
HOS	48,600	3,476	4,536
Sound category	5,602	290	716
Sound duration	7,754	310	1,076
Sound order	7,754	310	1,076
Total	69,710	4,386	7,404

Episodes are instructions with starting poses. Multiple episodes share a source clip or panorama; counts are not counts of independent videos. Original IDs, instructions, train/validation/test membership, and scoring are retained.

Contents

• code/: portable audio encoder, data preparation, alignment/SFT/AV-IDM, matched no-standalone-SFT ablation, evaluation, action parsing, and scoring.
• archives/: all indexed benchmark observations, panoramas, stereo audio, annotations, train/validation trajectories, referenced source videos and source metadata, and dry audio used by the auxiliary objective.
• metadata/files.jsonl: per-file sizes and SHA256 checksums.
• metadata/archives.json: archive sizes, SHA256 and member counts.
• scripts/extract_data.py: verify archives and extract matching data beside code.

The per-file manifest defines the exact release membership. Source metadata may mention upstream clips outside this benchmark; those unrelated clips are not recursively included. No pretrained model checkpoint is bundled.

Download and use

hf download zhizeng115/EchoScout-FourTask --repo-type dataset --local-dir EchoScout-FourTask
cd EchoScout-FourTask
python scripts/extract_data.py --destination code
cd code
python tests/check_code.py
python tests/check_numeric.py
python prepare_data.py merge
python prepare_data.py policy
python prepare_data.py auxiliary

Dataset paths in operational JSON/JSONL records are relative to the code/ working directory after extraction. Original upstream metadata is retained as provenance and may contain original source paths; it is not policy input. See code/README.md for dependencies, original model and audio-encoder inputs, training commands, checkpoint selection, and evaluation commands.

Evaluation protocol

The agent sees a 640×360, 100-degree horizontal-FOV view and at most two prior views/actions. HOS is visual-only. The three sound tasks provide the complete fixed 10-second recording before searching the frozen scene. The microphone orientation is fixed; camera rotations do not re-render binaural audio. Greedy generation allows up to 512 new tokens. Invalid actions consume a step without moving the camera. Submission must name the current pose. Frozen annotation-based geometric scorers are included; they are not pixel-mask IoU metrics. Test IDs are fixed and are not used for checkpoint selection.

Data quality and provenance

Labels are metadata-derived silver annotations, not fully human-verified gold labels. Sound duration and onset targets use scheduled event metadata. Scene split isolation is checked. Asset-level and pretrained-model exposure is not established; within-split identical panoramas may occur. Canonical per-channel RMS audio preprocessing removes absolute inter-channel level differences. Auxiliary dry mixtures duplicated across channels do not simulate binaural spatial acoustics.

Third-party code attribution and its MIT notice are retained in code/THIRD_PARTY_NOTICES.md and code/licenses/. Upstream media and pretrained models retain their source terms; this release does not assert a new blanket license over third-party material. This repository does not contain trained weights, credentials, private keys, remote-service configuration, or run logs.