code/DATA_FORMAT.md

External data layout

No records, identifiers, labels, audio, images, or videos are distributed here. Paths below are input/output contracts for running the source with separately provided assets. Paths stored inside JSONL records must resolve on the machine running the code. JSONL means one JSON object per line.

Raw source trajectories

datasets/source_trajectories/{static,sound}/{train,validation}.jsonl:

Each row contains id, episode_id, step, task, split, scene, clip_id, instruction, pose (integer yaw and pitch), panorama, and answer. Sound rows also contain audio_path. Static rows use source: "custom" and task: "HOS". Sound task names are sound_class, duration, and order. The source answer is an integer rotate or current-pose submit action enclosed in one <answer>...</answer> block. Step values must be consecutive.

Benchmark identity and annotations

datasets/omini_360_benchmark_v0_1/:

• public/{train,validation,test}.jsonl: instruction records; test records include id, task_type, instruction, initial_pose.
• private/episode_map.jsonl: maps id to scene, clip_id, and task_id.
• private/static_labels/hos_{train,validation}.jsonl: task_id, split, yaw, pitch, and bbox for the HOS geometric expert.

datasets/hstar_full_test_v1/<task_id>/ contains exactly one panorama PNG/JPEG and an annotation.json list. The first annotation supplies yaw and pitch intervals used by the static evaluator.

datasets/omini_sound_search_v1/:

• public/{train,validation,test}.jsonl: id, task, instruction, initial_pose, with source identity fields as needed by preparation.
• private/labels.jsonl: id, scene, clip_id, split, panorama, audio_path, target_index, and candidates.
• Each candidate contains center: [yaw,pitch], pixel bbox: [x1,y1,x2,y2], noun, and a dry_audio path for constructing auxiliary mixtures.

The sound scorer uses 512x2048 panorama annotation coordinates. Audio must supply at least ten seconds of two-channel finite samples. Exact test counts are HOS=4536, sound_class=716, duration=1076, order=1076. Ground-truth labels are for geometric supervision and scoring; the policy message builder never accepts labels.

Prepared policy data

workspace/fourtask_v2/data/ contains train.jsonl, validation.jsonl, train_episodes.jsonl, validation_episodes.jsonl, and audit.json. These are generated by prepare_data.py; they are not included in the archive.

Policy rows contain id, episode_id, task, split, scene, clip_id, instruction, pose, observations, history_actions, audio_path, audio_end_s, reference, dynamic, answers, supervision_kind, and recovery_prefix_steps. An observation is a panorama path and a camera pose. Rows use one to three observations and at most two historical actions.

Episode rows contain id, task, split, scene, clip_id, instruction, initial_pose, panorama, audio_path, and label. The integrity manifest records SHA256 for policy and episode files and requires the scene/clip split-disjointness flag. Do not fabricate audit manifests to bypass these checks; use the builders with matching source records.

Auxiliary data

workspace/fourtask_v3_audio_realign/data/ contains aux_train.jsonl, aux_validation.jsonl, generated waveforms, and their audit manifests. Each row contains source identity, paired candidates, a temporal query, target_index, onset/duration metadata, and the emitted waveform SHA256. The builder requires at least 100 accepted training clips and 16 validation clips. It verifies emitted energy windows and rejects unsuitable recordings.