qwen3_omni_thinker_data

# Qwen3-Omni Thinker Speculative Conditions Version: `2026.08` Conditions: `1,973,939` Unique media objects: `380,706` This is a self-contained input-condition dataset for generating trajectories with the Qwen3-Omni-30B-A3B-Instruct Thinker. The `conditions` configuration contains prompts and ordered `media_id` references. The `media` configuration contains the corresponding binary media. Source URLs are provenance only and are never required to train or regenerate Thinker responses. ```python import json from datasets import load_dataset conditions = load_dataset("<organization>/<dataset>", "conditions", split="train") media = load_dataset("<organization>/<dataset>", "media", split="train") media_by_id = {row["media_id"]: row for row in media} first_condition_media = [ media_by_id[media_id] for media_id in json.loads(conditions[0]["media_ids_json"]) ] ``` Each media row contains `bytes`, `media_type`, `mime_type`, `file_name`, a SHA-256 digest, provenance, and license metadata. Text-only conditions have an empty media list. The two configurations are one release unit and must be mirrored together. This package is intended for internal research in a private, access-controlled repository. Source and license metadata are retained on each condition and media row.