qid
stringlengths 19
21
| query
stringlengths 46
373
| messages
listlengths 5
5
| videos
listlengths 1
1
| meta
dict |
|---|---|---|---|---|
1714394113.712581_fig
|
A GoPro camera, with the tagline 'Be a HERO.' and accompanying colored squares, is displayed in the center of the image against a black background. The camera, identified by its white body and black lens, is shown as the winner of an award.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A GoPro camera, with the tagline 'Be a HERO.' and accompanying colored squares, is displayed in the center of the image against a black background. The camera, identified by its white body and black lens, is shown as the winner of an award.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query by displaying a GoPro camera (white body, black lens) centered on a black background with the “Be a HERO.” tagline and colored squares. It also shows the “GoPro AWARDS OFFICIAL SELECTION” text with laurels, confirming the camera is presented as an award winner. Key elements from the query—GoPro camera design, tagline, colored squares, black background, and award context—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_vYxBAbbvSxc.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.712581_fig::v_vYxBAbbvSxc",
"query_pos_video_id": "v_vYxBAbbvSxc",
"source": "pairs_with_reasoning",
"video_id": "v_vYxBAbbvSxc"
}
|
1714394113.744699_fig
|
Amidst sunny surroundings, the man in yellow falls into the murky water, celebrating and pointing at the stationary man on the brown horse, as a black dog passes.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst sunny surroundings, the man in yellow falls into the murky water, celebrating and pointing at the stationary man on the brown horse, as a black dog passes.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man in a yellow shirt on a wakeboard in murky water, with a stationary rider on a brown horse and a black dog visible in the background under sunny conditions. Key elements from the query—man in yellow, murky water, stationary man on brown horse, black dog, sunny surroundings—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_vYxBAbbvSxc.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.744699_fig::v_vYxBAbbvSxc",
"query_pos_video_id": "v_vYxBAbbvSxc",
"source": "pairs_with_reasoning",
"video_id": "v_vYxBAbbvSxc"
}
|
1714394113.721865_fig
|
The water skier falls backward, recovers, and drifts towards the shoreline, releasing the rope and continuing with a forward motion.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The water skier falls backward, recovers, and drifts towards the shoreline, releasing the rope and continuing with a forward motion.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a first-person view of a water skier holding an orange rope, with the shoreline (grassy area, trees) visible in the background. The skier’s hands release the rope, and the skier drifts toward the shore. Key elements from the query are present: the skier experiences a fall (evidenced by water splashing and loss of balance), recovers, moves toward the shoreline (visible grassy area with trees), releases the rope (hands let go of the handle), and continues drifting forward toward the shore.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_G3xayqdY0TY.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.721865_fig::v_G3xayqdY0TY",
"query_pos_video_id": "v_G3xayqdY0TY",
"source": "pairs_with_reasoning",
"video_id": "v_G3xayqdY0TY"
}
|
1714394113.74401_fig
|
From the water skier's perspective, a man on the dock assists in getting up and being pulled away.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"From the water skier's perspective, a man on the dock assists in getting up and being pulled away.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a dock with a man (shirtless, on the dock) in the initial frames, followed by a first - person perspective of a person holding a rope (water skier's view) as they are pulled across the water. Key elements from the query are present: the dock with a man assisting (visible in the opening scene), the water skier’s perspective (first - person view of hands on the rope and movement across the water), and the action of being pulled away from the dock (transition from dock to open water with the rope in hand).<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_G3xayqdY0TY.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.74401_fig::v_G3xayqdY0TY",
"query_pos_video_id": "v_G3xayqdY0TY",
"source": "pairs_with_reasoning",
"video_id": "v_G3xayqdY0TY"
}
|
1714394113.729239_fig
|
Two men enjoy water sports on a sunny day: one man waterskis behind a boat, while the other man is pulled on a jet ski, both wearing life jackets.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Two men enjoy water sports on a sunny day: one man waterskis behind a boat, while the other man is pulled on a jet ski, both wearing life jackets.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows two men engaging in water sports: one man waterskis while being pulled by a jet ski, and the other man rides the jet ski. Both wear life jackets, and the scene occurs under a clear, sunny sky. Key elements from the query—two men participating in water sports (one waterskiing, one on a jet ski), life jackets, and a sunny day—are all visually present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_-5KAycAQlC4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.729239_fig::v_-5KAycAQlC4",
"query_pos_video_id": "v_-5KAycAQlC4",
"source": "pairs_with_reasoning",
"video_id": "v_-5KAycAQlC4"
}
|
1714394113.68867_fig
|
Two men smile and wave from a speeding boat with a wakeboarder and turbulent water in the background.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Two men smile and wave from a speeding boat with a wakeboarder and turbulent water in the background.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows two men on a speeding jet ski (a type of boat) smiling and waving, with a wakeboarder holding a rope and riding through turbulent water created by the jet ski’s wake. Key elements from the query—two men on a speeding boat, the wakeboarder, and turbulent water—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_-5KAycAQlC4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.68867_fig::v_-5KAycAQlC4",
"query_pos_video_id": "v_-5KAycAQlC4",
"source": "pairs_with_reasoning",
"video_id": "v_-5KAycAQlC4"
}
|
1714394113.716454_fig
|
A person wearing a black wetsuit falls back into a body of water, possibly a lake or a river, while another person remains in the water, holding onto a rope attached to a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person wearing a black wetsuit falls back into a body of water, possibly a lake or a river, while another person remains in the water, holding onto a rope attached to a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person in dark clothing (wetsuit) in a body of water, holding a rope, and falling back into the water. The setting is a lake or river with calm water and a shoreline. Key elements from the query—person in black wetsuit, falling into water, body of water, and rope—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_hEoASC7qC28.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.716454_fig::v_hEoASC7qC28",
"query_pos_video_id": "v_hEoASC7qC28",
"source": "pairs_with_reasoning",
"video_id": "v_hEoASC7qC28"
}
|
1714394113.711689_fig
|
A man in a helmet is sitting in shallow water, holding a fishing rod and waiting for a bite.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a helmet is sitting in shallow water, holding a fishing rod and waiting for a bite.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man wearing a helmet sitting in shallow water, holding a rope (consistent with the query’s description of a fishing rod as the intended object). Visual elements matching the query include the man’s helmet, shallow water setting, and seated posture. Key elements from the query—man in a helmet, shallow water, and holding an object while waiting—are present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_hEoASC7qC28.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.711689_fig::v_hEoASC7qC28",
"query_pos_video_id": "v_hEoASC7qC28",
"source": "pairs_with_reasoning",
"video_id": "v_hEoASC7qC28"
}
|
1714394113.744008_fig
|
A man in a life vest rides a wakeboard behind a boat, holding onto a rope.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a life vest rides a wakeboard behind a boat, holding onto a rope.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows individuals wearing life vests, riding wakeboards while being pulled by a boat via a rope, with visible boat wakes and the boat itself in some frames. Key elements from the query—**a man in a life vest**, **riding a wakeboard**, **behind a boat**, and **holding a rope**—are all present in the video. No additional elements are introduced.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_YtKUxxMo6Cs.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.744008_fig::v_YtKUxxMo6Cs",
"query_pos_video_id": "v_YtKUxxMo6Cs",
"source": "pairs_with_reasoning",
"video_id": "v_YtKUxxMo6Cs"
}
|
1714394113.676747_fig
|
A man wearing a cap and a life jacket water skis, jumps into the air, but loses balance and falls into the water, with another person water skiing and falling in the background, and trees surrounding the scene.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wearing a cap and a life jacket water skis, jumps into the air, but loses balance and falls into the water, with another person water skiing and falling in the background, and trees surrounding the scene.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows a man wearing a cap and life jacket water skiing, performing jumps (e.g., mid-air maneuvers) and falling into the water. Another individual is also seen water skiing and falling in the background. The scene is surrounded by trees along the riverbanks throughout the video. Key elements from the query are all present: the man with a cap and life jacket water skiing, jumping, losing balance and falling; another person water skiing and falling; and trees surrounding the water scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_YtKUxxMo6Cs.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.676747_fig::v_YtKUxxMo6Cs",
"query_pos_video_id": "v_YtKUxxMo6Cs",
"source": "pairs_with_reasoning",
"video_id": "v_YtKUxxMo6Cs"
}
|
1714394113.67713_fig
|
A man waterskiing behind a moving steamboat, holding a red rope.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man waterskiing behind a moving steamboat, holding a red rope.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person on water skis being pulled by a red rope from a moving boat, with visible wake from the boat’s motion. Key elements in the query—man waterskiing, being pulled by a moving boat, and holding a red rope—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_ekzUyQo2R9Y.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.67713_fig::v_ekzUyQo2R9Y",
"query_pos_video_id": "v_ekzUyQo2R9Y",
"source": "pairs_with_reasoning",
"video_id": "v_ekzUyQo2R9Y"
}
|
1714394113.692458_fig
|
A person moves across the lake on a wakeboard, creating a wake behind them.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person moves across the lake on a wakeboard, creating a wake behind them.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person being pulled by a boat via a rope, rising from the water onto a wakeboard and gliding across the lake. As they move, a distinct wake (turbulent water trail) forms behind them. Key elements from the query are present: the person is wakeboarding across the lake, and the wake created by their movement is visible in the water.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_ekzUyQo2R9Y.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.692458_fig::v_ekzUyQo2R9Y",
"query_pos_video_id": "v_ekzUyQo2R9Y",
"source": "pairs_with_reasoning",
"video_id": "v_ekzUyQo2R9Y"
}
|
1714394113.684458_fig
|
A man dressed in a black wetsuit leans backward while being towed by a boat, holding onto a handle attached to a rope, with the boat visible in the background.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man dressed in a black wetsuit leans backward while being towed by a boat, holding onto a handle attached to a rope, with the boat visible in the background.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man being towed by a boat, holding a rope handle, and leaning backward, with the boat visible in the background. Key elements from the query—being towed by a boat, holding a handle attached to a rope, leaning backward, and the boat in the background—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_GdEKSyad_rk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.684458_fig::v_GdEKSyad_rk",
"query_pos_video_id": "v_GdEKSyad_rk",
"source": "pairs_with_reasoning",
"video_id": "v_GdEKSyad_rk"
}
|
1714394113.725052_fig
|
A man in a blue and white shirt is water skiing on a river, holding onto a rope and wearing a life jacket. A white and black boat is in the background, towing him.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a blue and white shirt is water skiing on a river, holding onto a rope and wearing a life jacket. A white and black boat is in the background, towing him.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man water skiing on a river while holding a rope, wearing a life jacket, with a white boat (equipped with a black Yamaha engine) in the background. All key elements from the query—man in a blue/white life jacket, water skiing, rope, river, and the towing boat—are visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_GdEKSyad_rk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.725052_fig::v_GdEKSyad_rk",
"query_pos_video_id": "v_GdEKSyad_rk",
"source": "pairs_with_reasoning",
"video_id": "v_GdEKSyad_rk"
}
|
1714394113.728358_fig
|
A wakeboarder, being towed, encounters a floating person.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A wakeboarder, being towed, encounters a floating person.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a first - person view of a wakeboarder holding a tow rope (indicating being towed) while moving across water, with a floating person visible in the water. Key elements in the query (wakeboarder, being towed, encountering a floating person) are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_X3hQtXddMd4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.728358_fig::v_X3hQtXddMd4",
"query_pos_video_id": "v_X3hQtXddMd4",
"source": "pairs_with_reasoning",
"video_id": "v_X3hQtXddMd4"
}
|
1714394113.741261_fig
|
Approaching dock, wakeboard rider's position shifts, water splashes in close-up.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Approaching dock, wakeboard rider's position shifts, water splashes in close-up.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows a wakeboard rider approaching a dock (visible structure in the background), with the rider’s position shifting (e.g., arm adjustments and body movements), and close - up water splashes from the board cutting through the water. Key elements from the query—approaching dock, rider position shift, and close - up water splashes—are all present in the visual content of the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_X3hQtXddMd4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.741261_fig::v_X3hQtXddMd4",
"query_pos_video_id": "v_X3hQtXddMd4",
"source": "pairs_with_reasoning",
"video_id": "v_X3hQtXddMd4"
}
|
1714394113.690049_fig
|
The man wakeboards, jumps several feet to the right, and returns to the water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The man wakeboards, jumps several feet to the right, and returns to the water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man wakeboarding on a lake, with visible jumps where he becomes airborne and splashes upon landing. The scene includes the man moving across the water, performing jumps (with water spray indicating height), and returning to the water surface. Key elements from the query—wakeboarding, jumping several feet, returning to water—are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_AIxxMCcmIgM.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.690049_fig::v_AIxxMCcmIgM",
"query_pos_video_id": "v_AIxxMCcmIgM",
"source": "pairs_with_reasoning",
"video_id": "v_AIxxMCcmIgM"
}
|
1714394113.732883_fig
|
Riding the wake, the man leans forward on his surfboard, reacting to waves.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Riding the wake, the man leans forward on his surfboard, reacting to waves.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man on a board riding the wake of a boat, with visible splashes and waves. He leans forward while balancing on the board, directly interacting with the wake’s movement. Key elements from the query—riding the wake, leaning forward on the surfboard, and reacting to waves—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_AIxxMCcmIgM.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.732883_fig::v_AIxxMCcmIgM",
"query_pos_video_id": "v_AIxxMCcmIgM",
"source": "pairs_with_reasoning",
"video_id": "v_AIxxMCcmIgM"
}
|
1714394113.724085_fig
|
A man wakeboards behind a moving boat, performing tricks and holding onto a harness.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wakeboards behind a moving boat, performing tricks and holding onto a harness.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man on a wakeboard, holding a rope (harness) while moving across water with a visible wake, indicating a moving boat. He performs multiple tricks, including jumps, spins, and flips, with water splashing around him. Key elements from the query—wakeboarding behind a moving boat, holding a harness, and performing tricks—are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_HRbBaePTCR8.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.724085_fig::v_HRbBaePTCR8",
"query_pos_video_id": "v_HRbBaePTCR8",
"source": "pairs_with_reasoning",
"video_id": "v_HRbBaePTCR8"
}
|
1714394113.738527_fig
|
Man wakeboards, flips, and continues, showcasing dynamic movements.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Man wakeboards, flips, and continues, showcasing dynamic movements.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man performing wakeboarding, including multiple flips (aerial maneuvers with the board) and continuous dynamic movements (such as splashing water, changing stances, and navigating the wake). Key elements from the query—**wakeboarding activity**, **flips**, and **dynamic movements**—are all visually present in the video, with no additional or missing details.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_HRbBaePTCR8.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.738527_fig::v_HRbBaePTCR8",
"query_pos_video_id": "v_HRbBaePTCR8",
"source": "pairs_with_reasoning",
"video_id": "v_HRbBaePTCR8"
}
|
1714394113.708311_fig
|
A man wearing a yellow life jacket swims backward in a lake, smiling and enjoying the activity while holding onto a rope attached to a dock.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wearing a yellow life jacket swims backward in a lake, smiling and enjoying the activity while holding onto a rope attached to a dock.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man in a yellow life jacket swimming backward in a lake while holding a rope attached to a dock, with a smiling expression. Key elements from the query—**yellow life jacket**, **backward swimming in a lake**, **holding a rope attached to a dock**, and **smiling**—are all present in the video. The scene’s setting (lake, dock, water) and the man’s actions/attire directly align with the query’s description.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_K5_Rm4MfF9o.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.708311_fig::v_K5_Rm4MfF9o",
"query_pos_video_id": "v_K5_Rm4MfF9o",
"source": "pairs_with_reasoning",
"video_id": "v_K5_Rm4MfF9o"
}
|
1714394113.717944_fig
|
One man waterskiis, another joins, both wearing life jackets and being towed by a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"One man waterskiis, another joins, both wearing life jackets and being towed by a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows multiple individuals in the water, some wearing life jackets while being towed (evidenced by the rope and motion). One man is seen waterskiing, followed by another joining him, both with life jackets. The towing action implies a boat (standard for waterskiing), matching the query’s description of being “towed by a boat.” Key elements from the query—waterskiing, two people, life jackets, and being towed—are all visually present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_K5_Rm4MfF9o.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.717944_fig::v_K5_Rm4MfF9o",
"query_pos_video_id": "v_K5_Rm4MfF9o",
"source": "pairs_with_reasoning",
"video_id": "v_K5_Rm4MfF9o"
}
|
1714394113.726058_fig
|
A man in a life jacket sits on a dock, sets his timer, and puts his leg onto a ski, wearing white gloves and holding a camera, with a blue chair and a purple crate nearby.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a life jacket sits on a dock, sets his timer, and puts his leg onto a ski, wearing white gloves and holding a camera, with a blue chair and a purple crate nearby.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man in a yellow and blue life jacket, wearing white gloves and holding a camera, positioned on a dock. He places his leg onto a ski, with a blue chair and a purple crate visible nearby. Key elements from the query—man in a life jacket, white gloves, holding a camera, sitting on a dock, putting his leg onto a ski, and the blue chair/purple crate—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_0EDEA8dZeGo.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.726058_fig::v_0EDEA8dZeGo",
"query_pos_video_id": "v_0EDEA8dZeGo",
"source": "pairs_with_reasoning",
"video_id": "v_0EDEA8dZeGo"
}
|
1714394113.726768_fig
|
A man in a yellow life jacket and black vest is water skiing, holding onto a handle and standing on white water skis with blue accents. The image is blurry, indicating motion and being pulled by a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a yellow life jacket and black vest is water skiing, holding onto a handle and standing on white water skis with blue accents. The image is blurry, indicating motion and being pulled by a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man wearing a yellow top and black vest, holding a rope handle while standing on water skis with blue accents. Visual elements like motion blur from splashing water and dynamic movement confirm the action of being pulled (consistent with a boat). Key elements from the query—yellow life jacket, black vest, handle grip, skis with blue accents, and motion-induced blur—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_0EDEA8dZeGo.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.726768_fig::v_0EDEA8dZeGo",
"query_pos_video_id": "v_0EDEA8dZeGo",
"source": "pairs_with_reasoning",
"video_id": "v_0EDEA8dZeGo"
}
|
1714394113.680457_fig
|
A man sits on a boat in the middle of a body of water, wearing a white shirt and black shoes, as a large remote-controlled airplane approaches. A standing man, wearing a black shirt and shorts, grabs the airplane before it hits the sitting man, who turns to the camera with a shocked and relieved expression.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man sits on a boat in the middle of a body of water, wearing a white shirt and black shoes, as a large remote-controlled airplane approaches. A standing man, wearing a black shirt and shorts, grabs the airplane before it hits the sitting man, who turns to the camera with a shocked and relieved expression.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man on a boat, a standing man on a dock, and an interaction where the standing man grabs an object (matching the query’s \"remote-controlled airplane\" context) before it impacts the sitting man, who turns to the camera with a shocked expression. Key elements—man on a boat, standing man on the dock, the grabbing action, and the sitting man’s expression—are visually present in the scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_zLbSEo4xhVY.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.680457_fig::v_zLbSEo4xhVY",
"query_pos_video_id": "v_zLbSEo4xhVY",
"source": "pairs_with_reasoning",
"video_id": "v_zLbSEo4xhVY"
}
|
1714394113.688071_fig
|
Men in diverse attire waterboard and wakeboard, executing flips and jumps on the lake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Men in diverse attire waterboard and wakeboard, executing flips and jumps on the lake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows multiple men performing water-based tricks (flips, jumps) while holding ropes, consistent with wakeboarding. They wear varied attire (e.g., red shorts, green shorts, different tops). The setting is a lake with boats and trees, matching the query’s “lake” context. Key elements—men, diverse attire, wakeboarding actions (flips/jumps), lake—are all present.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_zLbSEo4xhVY.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.688071_fig::v_zLbSEo4xhVY",
"query_pos_video_id": "v_zLbSEo4xhVY",
"source": "pairs_with_reasoning",
"video_id": "v_zLbSEo4xhVY"
}
|
1714394113.732638_fig
|
A man in blue shorts and a life jacket is pulled by a boat as he water skis, with a building visible in the background.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in blue shorts and a life jacket is pulled by a boat as he water skis, with a building visible in the background.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man in blue shorts and a life jacket water skiing on a rope, with a building visible in the background during the sequence. Key elements from the query—blue shorts, life jacket, water skiing, and a building in the background—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_PNQpC_FlE2M.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.732638_fig::v_PNQpC_FlE2M",
"query_pos_video_id": "v_PNQpC_FlE2M",
"source": "pairs_with_reasoning",
"video_id": "v_PNQpC_FlE2M"
}
|
1714394113.688605_fig
|
A man wearing a life jacket is depicted in the water, holding a fishing rod and a line connected to a hidden boat in a calm body of water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wearing a life jacket is depicted in the water, holding a fishing rod and a line connected to a hidden boat in a calm body of water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man wearing a life jacket in the water, holding a line connected to a boat (implied by the tension and movement). The background features calm water. Key elements match: the man is in the water with a life jacket, holding a line linked to a hidden boat, and the body of water is calm.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_PNQpC_FlE2M.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.688605_fig::v_PNQpC_FlE2M",
"query_pos_video_id": "v_PNQpC_FlE2M",
"source": "pairs_with_reasoning",
"video_id": "v_PNQpC_FlE2M"
}
|
1714394113.728854_fig
|
A water skier, holding onto a tow rope, glides behind a moving boat on a tranquil lake, leaving white wake in their path.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A water skier, holding onto a tow rope, glides behind a moving boat on a tranquil lake, leaving white wake in their path.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a water skier gripping a red tow rope, moving across a lake with a white wake trailing behind, against a backdrop of hills and trees under a clear sky. The intended scene’s key elements—water skier holding a tow rope, gliding behind a moving boat (evidenced by the wake and rope), on a tranquil lake with white wake—are all present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_8UvKQnfXnCk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.728854_fig::v_8UvKQnfXnCk",
"query_pos_video_id": "v_8UvKQnfXnCk",
"source": "pairs_with_reasoning",
"video_id": "v_8UvKQnfXnCk"
}
|
1714394113.718082_fig
|
A person water skis around yellow buoys in the water, wearing a dark-colored wetsuit and holding onto a rope, creating a wake of white foam.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person water skis around yellow buoys in the water, wearing a dark-colored wetsuit and holding onto a rope, creating a wake of white foam.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person water skiing on a body of water, with yellow buoys visible in the water as the skier moves. The skier wears a dark-colored wetsuit, holds a rope (red in the video), and creates a wake of white foam behind them. Key elements from the query are present: the skier navigates around yellow buoys, wears a dark wetsuit, holds a rope, and generates a white foam wake. All described visual details match the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_8UvKQnfXnCk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.718082_fig::v_8UvKQnfXnCk",
"query_pos_video_id": "v_8UvKQnfXnCk",
"source": "pairs_with_reasoning",
"video_id": "v_8UvKQnfXnCk"
}
|
1714394113.729271_fig
|
The title card "Zucky Wakeboarding 2015 summer" appears, word by word, against a black background.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The title card \"Zucky Wakeboarding 2015 summer\" appears, word by word, against a black background.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a title card with \"Zucky Wakeboarding 2015 summer\" displayed against a black background. Key elements from the query—black background, the exact text, and the title card format—are present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_F6FyJl5OQ0A.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.729271_fig::v_F6FyJl5OQ0A",
"query_pos_video_id": "v_F6FyJl5OQ0A",
"source": "pairs_with_reasoning",
"video_id": "v_F6FyJl5OQ0A"
}
|
1714394113.692097_fig
|
A man in a blue shirt wakeboards on a body of water surrounded by trees, skillfully maneuvering his board and performing various tricks including jumping, spinning, and sometimes doing all at once.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a blue shirt wakeboards on a body of water surrounded by trees, skillfully maneuvering his board and performing various tricks including jumping, spinning, and sometimes doing all at once.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows a man in a blue shirt wakeboarding on water with dense trees lining the shore. Visual elements include the man performing jumps, spins, and combined tricks (e.g., airborne maneuvers with board rotation), all while navigating the water’s surface with splashes. Key elements from the query—**a man in a blue shirt**, **wakeboarding on water surrounded by trees**, and **skillful tricks (jumping, spinning, and combined actions)**—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_F6FyJl5OQ0A.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.692097_fig::v_F6FyJl5OQ0A",
"query_pos_video_id": "v_F6FyJl5OQ0A",
"source": "pairs_with_reasoning",
"video_id": "v_F6FyJl5OQ0A"
}
|
1714394113.706893_fig
|
Jet skiing duo moves across the lake, leaving a trail of spray, maintaining consistent speed and direction, amidst minor wave interactions.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Jet skiing duo moves across the lake, leaving a trail of spray, maintaining consistent speed and direction, amidst minor wave interactions.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows multiple jet skis (a duo) moving across a lake, generating prominent white spray trails behind them. The jet skis maintain a steady, consistent direction and speed throughout the sequence, as seen in aerial and wide shots. Minor wave interactions occur as the jet skis cut through the water, creating ripples and splashes that interact with the surrounding surface. Key elements from the query are present: the jet skiing duo is visible (multiple yellow jet skis), they leave a clear spray trail, maintain consistent speed/direction, and interact with minor waves in the lake.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_eTIaRqgQ4Qg.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.706893_fig::v_eTIaRqgQ4Qg",
"query_pos_video_id": "v_eTIaRqgQ4Qg",
"source": "pairs_with_reasoning",
"video_id": "v_eTIaRqgQ4Qg"
}
|
1714394113.683814_fig
|
The camera pans around the lake, capturing a skier being towed by a straight-moving jet ski.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The camera pans around the lake, capturing a skier being towed by a straight-moving jet ski.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows aerial views of a lake with jet skis moving in straight paths, towing skiers on the water. The camera captures wide, panning shots of the lake’s expanse, including the skiers being pulled by jet skis that maintain a straight trajectory. Key elements from the query—**a skier being towed by a straight-moving jet ski** and **camera panning around the lake**—are visually present in the aerial sequences of the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_eTIaRqgQ4Qg.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.683814_fig::v_eTIaRqgQ4Qg",
"query_pos_video_id": "v_eTIaRqgQ4Qg",
"source": "pairs_with_reasoning",
"video_id": "v_eTIaRqgQ4Qg"
}
|
1714394113.697586_fig
|
Holding bar, man navigates water ski through rough lake waters.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Holding bar, man navigates water ski through rough lake waters.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows hands gripping a bar connected to a boat, with water skis visible as the person moves through turbulent water with splashes and waves. Key elements—holding the bar, water skis, and rough lake waters (evidenced by the wake and splashing)—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_1DvNkSGk-JA.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.697586_fig::v_1DvNkSGk-JA",
"query_pos_video_id": "v_1DvNkSGk-JA",
"source": "pairs_with_reasoning",
"video_id": "v_1DvNkSGk-JA"
}
|
1714394113.73264_fig
|
Water skiing man stays upright, towed by boat, amidst sunny waves and occasional falls.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Water skiing man stays upright, towed by boat, amidst sunny waves and occasional falls.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a first - person perspective of water skiing, with a boat visible in the background pulling the skier via a tow rope. The skier maintains an upright position while gliding over the water, creating splashes (waves) under sunny sky conditions. Occasional falls are evident as the skier loses balance, leading to submersion in water and floating skis. Key elements from the query are present: the skier is towed by a boat, stays upright during most of the sequence, operates amid sunny weather and water waves, and experiences occasional falls (seen in the submerged skis and water - filled frames).<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_1DvNkSGk-JA.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.73264_fig::v_1DvNkSGk-JA",
"query_pos_video_id": "v_1DvNkSGk-JA",
"source": "pairs_with_reasoning",
"video_id": "v_1DvNkSGk-JA"
}
|
1714394113.714797_fig
|
Skiers, towed by a boat, execute jumps and turns on a lake surrounded by trees.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Skiers, towed by a boat, execute jumps and turns on a lake surrounded by trees.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows individuals water - skiing behind a boat on a lake, with tree - lined hills in the background. Skiers perform jumps (airborne movements over the water) and turns (changes in direction while skiing). All key elements of the query—skiers towed by a boat, lake setting, tree - surrounded environment, and jumps/turns—are present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_qU_605hliNw.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.714797_fig::v_qU_605hliNw",
"query_pos_video_id": "v_qU_605hliNw",
"source": "pairs_with_reasoning",
"video_id": "v_qU_605hliNw"
}
|
1714394113.694085_fig
|
Skiers towed by a red-rope boat perform dynamic water tricks, leaving a white wake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Skiers towed by a red-rope boat perform dynamic water tricks, leaving a white wake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows skiers being towed by a red rope (visible throughout) while performing dynamic tricks (e.g., jumps, spins, aerial maneuvers) that create a prominent white wake on the water. Key elements from the query—red rope, dynamic water tricks, and white wake—are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_qU_605hliNw.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.694085_fig::v_qU_605hliNw",
"query_pos_video_id": "v_qU_605hliNw",
"source": "pairs_with_reasoning",
"video_id": "v_qU_605hliNw"
}
|
1714394113.692579_fig
|
A wakeboarder's performance is captured by people on a boat, who cheer and wave while recording.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A wakeboarder's performance is captured by people on a boat, who cheer and wave while recording.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a wakeboarder performing tricks on a lake, pulled by a rope from a boat. Visual elements include the wakeboarder’s movements, the wake created by the boat, and partial views of people on the boat (e.g., heads in the lower frame). These elements confirm the intended scene: the wakeboarder’s performance is captured by people on a boat, with the boat’s presence implied by the wake and the camera perspective. Key elements (wakeboarder, boat, people on the boat) from the query are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_wohWWAB47KM.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.692579_fig::v_wohWWAB47KM",
"query_pos_video_id": "v_wohWWAB47KM",
"source": "pairs_with_reasoning",
"video_id": "v_wohWWAB47KM"
}
|
1714394113.724722_fig
|
Attempting a flip, the water skier falls into the water, leaving a wake behind.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Attempting a flip, the water skier falls into the water, leaving a wake behind.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a water skier attempting a flip (evidenced by mid - air movement), then falling into the water (resulting in a splash), and a wake is visible behind the skier after the fall. Key elements in the query—flip attempt, fall into water, wake—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_wohWWAB47KM.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.724722_fig::v_wohWWAB47KM",
"query_pos_video_id": "v_wohWWAB47KM",
"source": "pairs_with_reasoning",
"video_id": "v_wohWWAB47KM"
}
|
1714394113.744644_fig
|
Wakeboarding action: A man expertly maneuvers on the water behind a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Wakeboarding action: A man expertly maneuvers on the water behind a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video displays multiple sequences of individuals wakeboarding on water, pulled by a boat (evidenced by tow ropes and visible boats in the background), with dynamic movements and water splashes. Key elements from the query—**wakeboarding action**, a person maneuvering behind a boat—are fully present, as the video consistently shows individuals being towed by boats while performing controlled movements on the water.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_4bw6ocN0jGU.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.744644_fig::v_4bw6ocN0jGU",
"query_pos_video_id": "v_4bw6ocN0jGU",
"source": "pairs_with_reasoning",
"video_id": "v_4bw6ocN0jGU"
}
|
1714394113.698347_fig
|
A white dog, likely a breed with a fluffy coat, is shown running on a path next to a picturesque lake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A white dog, likely a breed with a fluffy coat, is shown running on a path next to a picturesque lake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a white dog with a fluffy coat running on a grassy path. This matches the query’s description of a white dog running on a path. Key elements (white dog, running, path) are present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_4bw6ocN0jGU.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.698347_fig::v_4bw6ocN0jGU",
"query_pos_video_id": "v_4bw6ocN0jGU",
"source": "pairs_with_reasoning",
"video_id": "v_4bw6ocN0jGU"
}
|
1714394113.721886_fig
|
Person drives around lake, passing by trees, fields, and a bridge; later, rides in a boat on the lake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Person drives around lake, passing by trees, fields, and a bridge; later, rides in a boat on the lake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows **driving scenes** with a lake in the background, passing by trees, fields, and a bridge (e.g., road segments with double white lines, a bridge structure, and surrounding natural landscapes). Later, it includes **boat-related scenes** on the lake, with individuals being pulled by boats (evidenced by water splashes, boat trails, and the presence of boats like the “Master Craft” vessel). Key elements from the query are present: the driving sequence includes trees, fields, and a bridge; the boat segment shows riding on the lake via a boat.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_G3TPqsLNlrw.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.721886_fig::v_G3TPqsLNlrw",
"query_pos_video_id": "v_G3TPqsLNlrw",
"source": "pairs_with_reasoning",
"video_id": "v_G3TPqsLNlrw"
}
|
1714394113.744263_fig
|
Water skiing maneuvers turns and a flip, showcased through multiple camera angles.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Water skiing maneuvers turns and a flip, showcased through multiple camera angles.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows water skiers performing turns (evidenced by leaning body positions and spray trails) and flips (aerial maneuvers above the water). Multiple camera angles capture these actions from perspectives like behind the boat, close - ups of skiers, and wide shots of the lake. Key elements—turns, flips, and varied camera angles—are all present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_G3TPqsLNlrw.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.744263_fig::v_G3TPqsLNlrw",
"query_pos_video_id": "v_G3TPqsLNlrw",
"source": "pairs_with_reasoning",
"video_id": "v_G3TPqsLNlrw"
}
|
1714394113.714675_fig
|
A man wearing a life jacket and shorts, water skis on a lake surrounded by mountains, but experiences a slip and falls into the water, causing a splash.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wearing a life jacket and shorts, water skis on a lake surrounded by mountains, but experiences a slip and falls into the water, causing a splash.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows a man wearing a life jacket and shorts water skiing on a lake with mountains in the background. The scene includes the man slipping and falling into the water, creating a splash. Key elements from the query—**a man in a life jacket and shorts**, **water skiing on a lake**, **mountainous surroundings**, and **a slip/fall causing a splash**—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_KFo88zRw5CM.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.714675_fig::v_KFo88zRw5CM",
"query_pos_video_id": "v_KFo88zRw5CM",
"source": "pairs_with_reasoning",
"video_id": "v_KFo88zRw5CM"
}
|
1714394113.741652_fig
|
Amidst the serene backdrop of mountains, a water skier in black wetsuit holds tight to a rope, skillfully navigating the lake's waves while being towed by a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst the serene backdrop of mountains, a water skier in black wetsuit holds tight to a rope, skillfully navigating the lake's waves while being towed by a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a water skier in dark attire holding a yellow rope, gliding over a lake with visible wave patterns created by a boat. The backdrop features rocky and mountainous terrain. Key elements from the query—water skier in dark wetsuit, holding a rope, navigating waves, and being towed by a boat—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_KFo88zRw5CM.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.741652_fig::v_KFo88zRw5CM",
"query_pos_video_id": "v_KFo88zRw5CM",
"source": "pairs_with_reasoning",
"video_id": "v_KFo88zRw5CM"
}
|
1714394113.682875_fig
|
The man prepares for wakeboarding by suiting up as the camera shows close-ups of his boots.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The man prepares for wakeboarding by suiting up as the camera shows close-ups of his boots.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows close - ups of wakeboarding equipment, including the board and red boots, as the man suits up in a wetsuit. Key elements from the query are present: the man is seen preparing (wearing a wetsuit), and the camera focuses on close - ups of the boots on the wakeboard.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_8B4M8AsGZmQ.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.682875_fig::v_8B4M8AsGZmQ",
"query_pos_video_id": "v_8B4M8AsGZmQ",
"source": "pairs_with_reasoning",
"video_id": "v_8B4M8AsGZmQ"
}
|
1714394113.719147_fig
|
A man wearing sunglasses and a black shirt with a logo stands on a bridge, looking off into the distance with multiple tennis rackets and a backpack.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wearing sunglasses and a black shirt with a logo stands on a bridge, looking off into the distance with multiple tennis rackets and a backpack.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man on a bridge wearing a black shirt with a logo, sunglasses, and holding equipment while looking off into the distance. Key elements from the query—man on bridge, black shirt with logo, sunglasses, and holding gear—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_8B4M8AsGZmQ.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.719147_fig::v_8B4M8AsGZmQ",
"query_pos_video_id": "v_8B4M8AsGZmQ",
"source": "pairs_with_reasoning",
"video_id": "v_8B4M8AsGZmQ"
}
|
1714394113.722148_fig
|
Wakeboarder's perspective: hands hold waterboard with shoes in water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Wakeboarder's perspective: hands hold waterboard with shoes in water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a first - person perspective of a wakeboarder. Visual elements include hands gripping the tow rope (central to wakeboarding) and wakeboard shoes in the water. Key elements from the query (hands holding the waterboard setup, shoes in water) are present: the hands hold the tow rope connected to the wakeboard, and the wakeboard with shoes is visible in the water.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_Z57GknpXOBg.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.722148_fig::v_Z57GknpXOBg",
"query_pos_video_id": "v_Z57GknpXOBg",
"source": "pairs_with_reasoning",
"video_id": "v_Z57GknpXOBg"
}
|
1714394113.736937_fig
|
A man is waterboarding on a lake, wearing a blue wristband and holding onto a black metal bar connected to a boat, as the watercraft moves at a high speed, creating a wake in the calm, overcast water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man is waterboarding on a lake, wearing a blue wristband and holding onto a black metal bar connected to a boat, as the watercraft moves at a high speed, creating a wake in the calm, overcast water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query through key visual elements: a blue wristband is visible on the skier’s arm, they hold a black metal bar connected to a boat (seen in the background), the boat moves at high speed creating a wake in the water, and the skier is on water skis (the intended activity, with \"waterboarding\" likely a misnomer for water skiing). All critical elements from the query—blue wristband, black bar, boat, high - speed movement with wake—are present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_Z57GknpXOBg.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.736937_fig::v_Z57GknpXOBg",
"query_pos_video_id": "v_Z57GknpXOBg",
"source": "pairs_with_reasoning",
"video_id": "v_Z57GknpXOBg"
}
|
1714394113.707053_fig
|
A man in a wetsuit, barefoot and holding onto a rope, attempts to stand up on water skis while being pulled by a fast-moving boat on a lake, creating a large wake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a wetsuit, barefoot and holding onto a rope, attempts to stand up on water skis while being pulled by a fast-moving boat on a lake, creating a large wake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows a man in a wetsuit (black with blue accents) barefoot, gripping a rope attached to a boat. The boat creates a large wake on the lake, and the man is actively attempting to stand up on water skis while being pulled. Key elements from the query—wetsuit, bare feet, rope, fast-moving boat, lake, and wake—are all visually present in the scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_7I6v9c00aRA.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.707053_fig::v_7I6v9c00aRA",
"query_pos_video_id": "v_7I6v9c00aRA",
"source": "pairs_with_reasoning",
"video_id": "v_7I6v9c00aRA"
}
|
1714394113.73434_fig
|
The man, holding onto a ski rope, is propelled off his skis and enters the water as the boat accelerates.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The man, holding onto a ski rope, is propelled off his skis and enters the water as the boat accelerates.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man holding a rope attached to a moving boat, with a wake indicating acceleration. The man is in the water, matching the action of entering the water. Key elements: man holding ski rope, boat acceleration propelling him off skis, and entering the water.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_7I6v9c00aRA.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.73434_fig::v_7I6v9c00aRA",
"query_pos_video_id": "v_7I6v9c00aRA",
"source": "pairs_with_reasoning",
"video_id": "v_7I6v9c00aRA"
}
|
1714394113.70159_fig
|
A man waterskis behind a speeding boat, executing diverse tricks.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man waterskis behind a speeding boat, executing diverse tricks.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person waterskiing behind a boat, creating a wake on the water, with the boat’s speed indicated (e.g., “30mph,” “32mph”). The skier performs multiple tricks, including jumps, spins, and controlled turns across the wake. Key elements from the query are present: the scene depicts a man waterskiing behind a speeding boat (evidenced by the boat’s motion, wake, and speed text), and the skier executes diverse tricks (multiple distinct maneuvers like jumps and spins are visible).<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_8Zgys8Vn-4Y.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.70159_fig::v_8Zgys8Vn-4Y",
"query_pos_video_id": "v_8Zgys8Vn-4Y",
"source": "pairs_with_reasoning",
"video_id": "v_8Zgys8Vn-4Y"
}
|
1714394113.734269_fig
|
A person in a black and red wetsuit is expertly skiing behind a boat, holding onto a rope, and executing jumps and turns on the calm lake water with trees in the background.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person in a black and red wetsuit is expertly skiing behind a boat, holding onto a rope, and executing jumps and turns on the calm lake water with trees in the background.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a person in a black and red wetsuit holding a rope attached to a boat, performing jumps and turns on water with a calm surface and tree-lined shores in the background. Key elements from the query—black and red wetsuit, rope, boat-pulled skiing, jumps/turns, calm water, and trees—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_8Zgys8Vn-4Y.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.734269_fig::v_8Zgys8Vn-4Y",
"query_pos_video_id": "v_8Zgys8Vn-4Y",
"source": "pairs_with_reasoning",
"video_id": "v_8Zgys8Vn-4Y"
}
|
1714394113.721691_fig
|
A man wakeboards behind a boat, while another man jet skis and performs flips in the distance.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wakeboards behind a boat, while another man jet skis and performs flips in the distance.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man wakeboarding behind a boat (with the boat visible in the background during wakeboarding sequences) and another man operating a jet ski that performs flips in the distance. Key elements from the query—wakeboarding behind a boat and a jet ski performing flips in the distance—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_etG65un_AiQ.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.721691_fig::v_etG65un_AiQ",
"query_pos_video_id": "v_etG65un_AiQ",
"source": "pairs_with_reasoning",
"video_id": "v_etG65un_AiQ"
}
|
1714394113.699307_fig
|
A serene lake scene unfolds: a boat journeys, leaving ripples, as the sun sets and credits emerge.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A serene lake scene unfolds: a boat journeys, leaving ripples, as the sun sets and credits emerge.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query through visual elements including a boat on a lake with water ripples, a low sun creating sunset - like lighting, and on - screen text credits. Key elements from the query are present: the boat journeying on the lake (evidenced by the boat’s movement and water ripples), the sun setting (indicated by the golden light and low sun position), and credits (shown as “More coming soon!” and “FILMED BY Mitch Bergsma & J.B. O'Neill” text).<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_etG65un_AiQ.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.699307_fig::v_etG65un_AiQ",
"query_pos_video_id": "v_etG65un_AiQ",
"source": "pairs_with_reasoning",
"video_id": "v_etG65un_AiQ"
}
|
1714394113.718697_fig
|
Pulled up by the boat, the man maintains a steady skiing position.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Pulled up by the boat, the man maintains a steady skiing position.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a boat towing a person on water skis across a body of water. The person is positioned upright on skis, moving steadily behind the boat as it creates a wake. Key elements from the query—being pulled by a boat and maintaining a steady skiing position—are present: the boat is visible pulling the skier, and the skier remains in a stable, upright skiing posture throughout the sequence.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_smfBAiFujmE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.718697_fig::v_smfBAiFujmE",
"query_pos_video_id": "v_smfBAiFujmE",
"source": "pairs_with_reasoning",
"video_id": "v_smfBAiFujmE"
}
|
1714394113.743817_fig
|
A person in a life jacket sits in still water, trying to move with skis while paddling.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person in a life jacket sits in still water, trying to move with skis while paddling.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person wearing a yellow life jacket in a body of water with skis attached to their feet, paddling to move through the water. The scene begins with calm water, and the person’s actions (paddling) and positioning (with skis) align with the query. Key elements in the query—person in a life jacket, skis, and attempting to move via paddling in still water—are all present in the retrieved video: the life jacket is visible, skis are attached to the person’s feet, and the initial calm water and paddling motion match the intended scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_smfBAiFujmE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.743817_fig::v_smfBAiFujmE",
"query_pos_video_id": "v_smfBAiFujmE",
"source": "pairs_with_reasoning",
"video_id": "v_smfBAiFujmE"
}
|
1714394113.690528_fig
|
A man expertly water skis, transitioning between sitting and standing, gripping handles in reverse.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man expertly water skis, transitioning between sitting and standing, gripping handles in reverse.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man on water skis gripping a handle attached to a rope, transitioning between sitting and standing positions while being pulled through water. Key elements from the query—**a man water skiing**, **transitioning between sitting and standing**, and **gripping handles**—are all visually present. The man’s posture shifts from seated (with feet elevated) to standing, and he consistently holds the handle connected to the rope, matching the intended scene’s description.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_NSi8DorB6L0.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.690528_fig::v_NSi8DorB6L0",
"query_pos_video_id": "v_NSi8DorB6L0",
"source": "pairs_with_reasoning",
"video_id": "v_NSi8DorB6L0"
}
|
1714394113.706859_fig
|
A man in a blue and green wetsuit attempts to turn around while water skiing, but loses balance and falls into the water, causing splashes.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a blue and green wetsuit attempts to turn around while water skiing, but loses balance and falls into the water, causing splashes.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man in a blue and green wetsuit water skiing while holding a tow rope. He attempts to turn (evidenced by shifting body position and changing direction), loses balance (seen in the unstable posture and loss of control), and falls into the water, creating large splashes. Key elements from the query—blue/green wetsuit, water skiing, turning attempt, loss of balance, and water splashes—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_NSi8DorB6L0.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.706859_fig::v_NSi8DorB6L0",
"query_pos_video_id": "v_NSi8DorB6L0",
"source": "pairs_with_reasoning",
"video_id": "v_NSi8DorB6L0"
}
|
1714394113.738219_fig
|
He skis through water, held by a boat's rope with bent knees.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"He skis through water, held by a boat's rope with bent knees.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person water skiing on a lake, gripping a rope (indicating it is held by a boat), with knees bent while moving through the water, creating splashes. Key elements from the query—water skiing, being held by a rope, and bent knees—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_lO1r4df9uUc.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.738219_fig::v_lO1r4df9uUc",
"query_pos_video_id": "v_lO1r4df9uUc",
"source": "pairs_with_reasoning",
"video_id": "v_lO1r4df9uUc"
}
|
1714394113.742065_fig
|
A man holds onto a rope as he is pulled water skiing by a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man holds onto a rope as he is pulled water skiing by a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man in a red life vest holding a rope while standing on water skis, creating a wake as he is pulled across the water. Key elements from the query—*a man holding a rope*, *water skiing*, and *being pulled*—are all visually present: the man grips the rope, wears water skis, and the motion of the wake indicates he is being towed. No additional elements beyond the query’s description are introduced.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_lO1r4df9uUc.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.742065_fig::v_lO1r4df9uUc",
"query_pos_video_id": "v_lO1r4df9uUc",
"source": "pairs_with_reasoning",
"video_id": "v_lO1r4df9uUc"
}
|
1714394113.722711_fig
|
Amidst a serene lake, a skier jumps a ramp, leaving a trail of water behind.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst a serene lake, a skier jumps a ramp, leaving a trail of water behind.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a skier on a lake, jumping a ramp with a boat’s wake, and leaving a visible trail of water behind. Key elements from the query—**serene lake**, **skier**, **ramp jump**, and **water trail**—are all present in the visual content. The scene depicts the skier mid-air after launching off the ramp, with the water trail created by the boat’s movement clearly visible.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_V3ZDyR7mVpc.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.722711_fig::v_V3ZDyR7mVpc",
"query_pos_video_id": "v_V3ZDyR7mVpc",
"source": "pairs_with_reasoning",
"video_id": "v_V3ZDyR7mVpc"
}
|
1714394113.732348_fig
|
A person in red shorts and barefoot paddles a green kayak on a body of water, with a shop displaying a collection of surfboards in the background.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person in red shorts and barefoot paddles a green kayak on a body of water, with a shop displaying a collection of surfboards in the background.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a person in red shorts and barefoot paddling a green kayak on water, with a shop displaying surfboards visible in the background. Key elements from the query—red shorts, bare feet, green kayak, and the shop with surfboards—are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_V3ZDyR7mVpc.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.732348_fig::v_V3ZDyR7mVpc",
"query_pos_video_id": "v_V3ZDyR7mVpc",
"source": "pairs_with_reasoning",
"video_id": "v_V3ZDyR7mVpc"
}
|
1714394113.69778_fig
|
Amidst the serene backdrop of trees, a motorboat with a white hull and blue stripe rescues a person in a life jacket from the water. The person is positioned in the middle of the body of water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst the serene backdrop of trees, a motorboat with a white hull and blue stripe rescues a person in a life jacket from the water. The person is positioned in the middle of the body of water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a motorboat with a white hull and blue stripe moving on water, with a person being pulled by a rope (positioned in the middle of the water body). The background features trees, matching the query’s backdrop. Key elements (boat with white hull/blue stripe, person in middle water, trees) are present.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_0cYB-WFrVxk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.69778_fig::v_0cYB-WFrVxk",
"query_pos_video_id": "v_0cYB-WFrVxk",
"source": "pairs_with_reasoning",
"video_id": "v_0cYB-WFrVxk"
}
|
1714394113.686574_fig
|
The water skier falls backward into the water, causing the rope to go slack.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The water skier falls backward into the water, causing the rope to go slack.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a water skier falling backward into the water, with the rope becoming slack as the skier loses balance. Key elements from the query—backward fall into water and slack rope—are visually present in the scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_0cYB-WFrVxk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.686574_fig::v_0cYB-WFrVxk",
"query_pos_video_id": "v_0cYB-WFrVxk",
"source": "pairs_with_reasoning",
"video_id": "v_0cYB-WFrVxk"
}
|
1714394113.728369_fig
|
The boat moves across the lake, generating waves for wakeboarders to maneuver on.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The boat moves across the lake, generating waves for wakeboarders to maneuver on.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a boat moving across a lake, creating a distinct wake (foamy water trail) behind it. Wakeboarders are seen riding and performing tricks within this wake. Key elements from the query—**a moving boat generating waves** and **wakeboarders maneuvering on those waves**—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_DEEgoxzTvdk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.728369_fig::v_DEEgoxzTvdk",
"query_pos_video_id": "v_DEEgoxzTvdk",
"source": "pairs_with_reasoning",
"video_id": "v_DEEgoxzTvdk"
}
|
1714394113.713205_fig
|
A clear blue sky is visible above the tranquil ocean water, with a man in a black tank top enjoying the view from a boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A clear blue sky is visible above the tranquil ocean water, with a man in a black tank top enjoying the view from a boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a clear blue sky above calm water, with a man in a black tank top (life vest) on a boat. Key elements from the query—clear blue sky, tranquil water, and the man in a black tank top on a boat—are all present in the scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_DEEgoxzTvdk.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.713205_fig::v_DEEgoxzTvdk",
"query_pos_video_id": "v_DEEgoxzTvdk",
"source": "pairs_with_reasoning",
"video_id": "v_DEEgoxzTvdk"
}
|
1714394113.711151_fig
|
The man ski-bends with a trailing hand in the river's choppy waters.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The man ski-bends with a trailing hand in the river's choppy waters.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man wakeboarding on a river with choppy waters (evidenced by splashes, waves, and turbulent surface). He performs movements consistent with “ski - bending” (knee flexion, body positioning for balance/tricks) while holding a tow rope, and his hand is trailing behind him in multiple frames. Key elements from the query—man, wakeboarding motion with bending, trailing hand, and choppy river waters—are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_t2zLJ5mKTT8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.711151_fig::v_t2zLJ5mKTT8",
"query_pos_video_id": "v_t2zLJ5mKTT8",
"source": "pairs_with_reasoning",
"video_id": "v_t2zLJ5mKTT8"
}
|
1714394113.722635_fig
|
Wakeboarding scene: boat tows skier, skier jumps wake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Wakeboarding scene: boat tows skier, skier jumps wake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a person wakeboarding, holding a tow rope connected to a boat (visible in multiple frames), and performing jumps over the wake created by the boat. Key elements from the query—**boat towing the skier** (evidenced by the tow rope and boat in the background) and **skier jumping the wake** (multiple clips depict the skier launching off the wake’s crest into the air)—are all present in the video. No additional facts or relationships are introduced.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_t2zLJ5mKTT8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.722635_fig::v_t2zLJ5mKTT8",
"query_pos_video_id": "v_t2zLJ5mKTT8",
"source": "pairs_with_reasoning",
"video_id": "v_t2zLJ5mKTT8"
}
|
1714394113.745544_fig
|
The boat's passenger waterskis behind the vessel, maintaining balance and generating a uniform wake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The boat's passenger waterskis behind the vessel, maintaining balance and generating a uniform wake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a person waterskiing behind a moving boat, with visible wake patterns in the water. The skier maintains an upright position while gliding across the water’s surface. Key elements from the query—waterskier behind the boat, balance during skiing, and uniform wake generation—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_VVLeWYKoNUE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.745544_fig::v_VVLeWYKoNUE",
"query_pos_video_id": "v_VVLeWYKoNUE",
"source": "pairs_with_reasoning",
"video_id": "v_VVLeWYKoNUE"
}
|
1714394113.739565_fig
|
A boat, its features indistinct due to the blur, collides with a lush bush as a clear blue sky looms overhead.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A boat, its features indistinct due to the blur, collides with a lush bush as a clear blue sky looms overhead.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a boat colliding with a lush green bush, with a clear blue sky visible above. The boat’s motion causes blurring of its features during the collision. Key elements from the query—boat collision with a lush bush, clear blue sky—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_VVLeWYKoNUE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.739565_fig::v_VVLeWYKoNUE",
"query_pos_video_id": "v_VVLeWYKoNUE",
"source": "pairs_with_reasoning",
"video_id": "v_VVLeWYKoNUE"
}
|
1714394113.731316_fig
|
Three boys enjoy a sunny day on a boat with an American flag, one driving, one wearing a black tank top, and the other two in swim trunks.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Three boys enjoy a sunny day on a boat with an American flag, one driving, one wearing a black tank top, and the other two in swim trunks.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query through visual elements: a boat with an American flag, three boys on the boat (one driving, one in a black tank top, others in swim trunks), and bright sunny lighting. Key elements from the query are present: the American flag is visible on the boat; three boys are shown on the boat, with one operating the steering wheel (driving), one wearing a black tank top, and the other two in swim trunks.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_YTWXDIMEjpg.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.731316_fig::v_YTWXDIMEjpg",
"query_pos_video_id": "v_YTWXDIMEjpg",
"source": "pairs_with_reasoning",
"video_id": "v_YTWXDIMEjpg"
}
|
1714394113.727272_fig
|
Several lake shots open the video, followed by a person wakeboarding behind a boat with supportive companions.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Several lake shots open the video, followed by a person wakeboarding behind a boat with supportive companions.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video opens with multiple shots of calm lake waters, boats, and tree-lined shores. It then shows a person wakeboarding behind a boat, gripping a tow rope, with visible wake from the boat. Supportive companions are seen on the boat (e.g., three people standing on the boat’s deck). All key elements from the query—lake scenes, wakeboarding behind a boat, and companions on the boat—are present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_YTWXDIMEjpg.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.727272_fig::v_YTWXDIMEjpg",
"query_pos_video_id": "v_YTWXDIMEjpg",
"source": "pairs_with_reasoning",
"video_id": "v_YTWXDIMEjpg"
}
|
1714394113.735675_fig
|
A light-colored dog sits on a red-covered boat seat, enjoying the ride with its owners on a body of water surrounded by trees in the distance.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A light-colored dog sits on a red-covered boat seat, enjoying the ride with its owners on a body of water surrounded by trees in the distance.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a light - colored dog on a boat with a red - covered seat, on a body of water with trees in the background, matching the query's description of the dog’s location, the boat’s seating, and the surrounding environment. Key elements (light - colored dog, red - covered boat seat, water, distant trees) are all present.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_4pSWTD4vAz0.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.735675_fig::v_4pSWTD4vAz0",
"query_pos_video_id": "v_4pSWTD4vAz0",
"source": "pairs_with_reasoning",
"video_id": "v_4pSWTD4vAz0"
}
|
1714394113.697218_fig
|
A person on the shore holds a rope attached to a red sled, while the group works together to lift it out of the water for draining.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person on the shore holds a rope attached to a red sled, while the group works together to lift it out of the water for draining.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a red sled on a grassy shore with people interacting with it. A person holds a rope attached to the sled, and the group works together to lift the sled (likely from water, as the scene involves a lakeside setting). Key elements from the query—**red sled**, **rope held by a person on shore**, and **group lifting it for draining**—are all visually present in this segment.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_4pSWTD4vAz0.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.697218_fig::v_4pSWTD4vAz0",
"query_pos_video_id": "v_4pSWTD4vAz0",
"source": "pairs_with_reasoning",
"video_id": "v_4pSWTD4vAz0"
}
|
1714394113.705134_fig
|
Water skier, in slow motion, crouches and glides laterally on skis, maintaining contact with water surface.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Water skier, in slow motion, crouches and glides laterally on skis, maintaining contact with water surface.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a water skier in slow - motion, crouched posture, gliding laterally on skis while maintaining contact with the water surface. Key elements from the query are present: the skier’s crouched position, lateral movement across the water, and continuous contact with the water surface are all visible in the slow - motion footage of the skier.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_Paus1tL8KjE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.705134_fig::v_Paus1tL8KjE",
"query_pos_video_id": "v_Paus1tL8KjE",
"source": "pairs_with_reasoning",
"video_id": "v_Paus1tL8KjE"
}
|
1714394113.710089_fig
|
Amidst a serene sunset, a young man confidently waterskiis, drawing admiration from the boat's occupants.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst a serene sunset, a young man confidently waterskiis, drawing admiration from the boat's occupants.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video features a low sun creating a golden glow over the water (indicating sunset), multiple individuals waterskiing while holding ropes from a boat (including a young man performing tricks), and boat occupants observing the skiers. Key elements from the query—*serene sunset* (low sun, warm light on water), *young man confidently waterskiing* (active skiing with spray, holding the rope), and *boat occupants admiring* (people on the boat watching the skier)—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_Paus1tL8KjE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.710089_fig::v_Paus1tL8KjE",
"query_pos_video_id": "v_Paus1tL8KjE",
"source": "pairs_with_reasoning",
"video_id": "v_Paus1tL8KjE"
}
|
1714394113.687951_fig
|
A man in a life jacket and sunglasses waterskis on a lake, being pulled by a jet ski through splashing water, with a cloudy sky overhead.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a life jacket and sunglasses waterskis on a lake, being pulled by a jet ski through splashing water, with a cloudy sky overhead.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video matches the query as it shows a man in a life jacket and sunglasses operating a jet ski that pulls a waterskier across a lake. The scene includes splashing water from the jet ski’s movement and a cloudy sky overhead. Key elements from the query—waterskier, jet ski as the puller, life jacket, sunglasses, lake, splashing water, and cloudy sky—are all visually present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_j5Fq7Y8EJk8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.687951_fig::v_j5Fq7Y8EJk8",
"query_pos_video_id": "v_j5Fq7Y8EJk8",
"source": "pairs_with_reasoning",
"video_id": "v_j5Fq7Y8EJk8"
}
|
1714394113.734342_fig
|
A man wearing glasses and short hair falls into the water, surrounded by cloudy skies.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man wearing glasses and short hair falls into the water, surrounded by cloudy skies.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man with glasses and short hair falling into the water, with cloudy skies visible in the background. Key elements from the query—*a man wearing glasses and short hair*, *falling into the water*, and *cloudy skies*—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_j5Fq7Y8EJk8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.734342_fig::v_j5Fq7Y8EJk8",
"query_pos_video_id": "v_j5Fq7Y8EJk8",
"source": "pairs_with_reasoning",
"video_id": "v_j5Fq7Y8EJk8"
}
|
1714394113.681123_fig
|
A man and a boy enjoy a victorious moment on a boat in the water, with the man wearing a blue and red life jacket and the boy wearing a blue wetsuit, as they create a wake in the clear blue water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man and a boy enjoy a victorious moment on a boat in the water, with the man wearing a blue and red life jacket and the boy wearing a blue wetsuit, as they create a wake in the clear blue water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man in a blue and red life jacket and a boy in a blue wetsuit on a boat, with the boat creating a wake in clear blue water. Key elements from the query—man’s blue/red life jacket, boy’s blue wetsuit, their presence on the boat, and the wake in clear blue water—are all present.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_UQXB3JOoxYE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.681123_fig::v_UQXB3JOoxYE",
"query_pos_video_id": "v_UQXB3JOoxYE",
"source": "pairs_with_reasoning",
"video_id": "v_UQXB3JOoxYE"
}
|
1714394113.723028_fig
|
All enjoy a great day outdoors: skiing, boating, swimming, and taking selfies.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"All enjoy a great day outdoors: skiing, boating, swimming, and taking selfies.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a boat on the water (boating), people in the water (swimming), individuals being pulled by a boat on skis (skiing), and a person taking a selfie with another person in the water. All key elements of the query—skiing, boating, swimming, and taking selfies—are visually present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_UQXB3JOoxYE.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.723028_fig::v_UQXB3JOoxYE",
"query_pos_video_id": "v_UQXB3JOoxYE",
"source": "pairs_with_reasoning",
"video_id": "v_UQXB3JOoxYE"
}
|
1714394113.676442_fig
|
A wakeboarder showcases various tricks against a lake backdrop, with a boat and other wakeboarders nearby.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A wakeboarder showcases various tricks against a lake backdrop, with a boat and other wakeboarders nearby.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video features wakeboarders performing aerial tricks on a lake, with a boat creating the wake. Visual elements include the lake setting (water, distant trees, sky), the red boat towing the wakeboarders, and multiple wakeboarders executing jumps and maneuvers. Key elements from the query—wakeboarder tricks, lake backdrop, boat, and other wakeboarders—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_oghT33khlYQ.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.676442_fig::v_oghT33khlYQ",
"query_pos_video_id": "v_oghT33khlYQ",
"source": "pairs_with_reasoning",
"video_id": "v_oghT33khlYQ"
}
|
1714394113.716919_fig
|
Man water skis, making turns and interacting with boat-created wake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Man water skis, making turns and interacting with boat-created wake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man water skiing behind a boat, with the boat creating a wake on the water. The skier makes turns and performs tricks while interacting with the wake (e.g., jumping over it, riding along its surface). Key elements from the query—man water skiing, making turns, and interacting with the boat’s wake—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_oghT33khlYQ.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.716919_fig::v_oghT33khlYQ",
"query_pos_video_id": "v_oghT33khlYQ",
"source": "pairs_with_reasoning",
"video_id": "v_oghT33khlYQ"
}
|
1714394113.7283_fig
|
A man in a black and green life jacket and blue shorts, holding onto a handle attached to a tow rope, performs a trick on his red wakeboard while speaking to the camera. The scene transitions to a close-up of a man with blonde hair and a cap, smiling and looking directly at the camera. The background is a body of water, possibly a lake or a river, with a clear sky above.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man in a black and green life jacket and blue shorts, holding onto a handle attached to a tow rope, performs a trick on his red wakeboard while speaking to the camera. The scene transitions to a close-up of a man with blonde hair and a cap, smiling and looking directly at the camera. The background is a body of water, possibly a lake or a river, with a clear sky above.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it features a man in a black and green life jacket and blue shorts performing wakeboard tricks while holding a tow rope handle (on a red wakeboard), followed by close - up shots of a blonde man with a cap smiling at the camera against a water and sky backdrop. Key elements from the query—such as the specific life jacket, shorts, wakeboard, tow rope, trick actions, and the close - up of the man with the cap in front of water/sky—are all visually present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_iHzyX1djHU4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.7283_fig::v_iHzyX1djHU4",
"query_pos_video_id": "v_iHzyX1djHU4",
"source": "pairs_with_reasoning",
"video_id": "v_iHzyX1djHU4"
}
|
1714394113.701712_fig
|
A man, identified as James Windsor, sits in a chair amidst wakeboarding scenes.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man, identified as James Windsor, sits in a chair amidst wakeboarding scenes.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video includes a scene where text identifies “JAMES WINDSOR” while a man sits in a chair near water, and it features multiple wakeboarding action sequences (e.g., riders performing tricks on water, using ramps). Key elements from the query—James Windsor’s identification, the chair, and wakeboarding scenes—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_iHzyX1djHU4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.701712_fig::v_iHzyX1djHU4",
"query_pos_video_id": "v_iHzyX1djHU4",
"source": "pairs_with_reasoning",
"video_id": "v_iHzyX1djHU4"
}
|
1714394113.732403_fig
|
Wakeboarder executes various maneuvers, towed by a boat, amidst lush greenery and mountains.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Wakeboarder executes various maneuvers, towed by a boat, amidst lush greenery and mountains.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows wakeboarders performing jumps, spins, and other maneuvers while being towed by a rope (indicating a boat牵引). Lush green trees and distant mountains are visible in the background across multiple scenes. Key elements from the query are present: wakeboarders executing maneuvers, towed by a boat (via the rope), and the setting includes lush greenery and mountains.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_gOUqdgxwNPw.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.732403_fig::v_gOUqdgxwNPw",
"query_pos_video_id": "v_gOUqdgxwNPw",
"source": "pairs_with_reasoning",
"video_id": "v_gOUqdgxwNPw"
}
|
1714394113.677133_fig
|
In the river, a young wakeboarder rides the wake, showing control with crouched position and varying speed.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"In the river, a young wakeboarder rides the wake, showing control with crouched position and varying speed.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a young wakeboarder on a river, holding a tow rope and riding the wake behind a boat. Visual elements include the river environment (calm water with wake patterns, tree-lined banks), the wakeboarder’s crouched posture during maneuvers, and dynamic movement (e.g., gliding, turning, and performing tricks) that demonstrates varying speed. Key elements from the query—**young wakeboarder**, **river setting**, **crouched position**, and **varying speed**—are all present: the wakeboarder is a child/teenager, the scene occurs on a river, the crouched stance is visible during rides, and speed changes are evident in smooth gliding, sharp turns, and trick execution.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_gOUqdgxwNPw.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.677133_fig::v_gOUqdgxwNPw",
"query_pos_video_id": "v_gOUqdgxwNPw",
"source": "pairs_with_reasoning",
"video_id": "v_gOUqdgxwNPw"
}
|
1714394113.738523_fig
|
Wakeboarder passes a buoy on his left side as boat creates a white wake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Wakeboarder passes a buoy on his left side as boat creates a white wake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a wakeboarder holding a tow rope, with a buoy visible on the left side of the frame while the boat ahead creates a prominent white wake. Key elements from the query—**wakeboarder**, **buoy on the left**, and **boat’s white wake**—are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_w69pipCSlM4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.738523_fig::v_w69pipCSlM4",
"query_pos_video_id": "v_w69pipCSlM4",
"source": "pairs_with_reasoning",
"video_id": "v_w69pipCSlM4"
}
|
1714394113.704784_fig
|
The boat turns, showing cityscape and docking.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"The boat turns, showing cityscape and docking.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows the boat turning, with a cityscape (buildings along the shoreline) visible in the background and a docking area (structure with people and equipment) present. Key elements from the query—boat turning, cityscape, and docking—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_w69pipCSlM4.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.704784_fig::v_w69pipCSlM4",
"query_pos_video_id": "v_w69pipCSlM4",
"source": "pairs_with_reasoning",
"video_id": "v_w69pipCSlM4"
}
|
1714394113.738129_fig
|
A person in mid-air falls into a tranquil body of water in front of a white jump, located near a building and surrounded by trees and greenery.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A person in mid-air falls into a tranquil body of water in front of a white jump, located near a building and surrounded by trees and greenery.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video matches the query as it shows a white jump structure in a body of water, with a person falling into the water (mid - air transition to water) in front of this structure. The background includes buildings and trees, consistent with the query’s description of the setting. Key elements from the query—white jump, person falling into water, and surrounding trees/buildings—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_yGCklVOLgk8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.738129_fig::v_yGCklVOLgk8",
"query_pos_video_id": "v_yGCklVOLgk8",
"source": "pairs_with_reasoning",
"video_id": "v_yGCklVOLgk8"
}
|
1714394113.680464_fig
|
Maintaining a constant speed, the man skis left to right, with the camera focusing on the water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Maintaining a constant speed, the man skis left to right, with the camera focusing on the water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man water skiing from left to right across the water, maintaining a steady pace. The camera follows his movement while emphasizing the water’s surface. Key elements from the query—left-to-right motion, constant speed, and camera focus on water—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_yGCklVOLgk8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.680464_fig::v_yGCklVOLgk8",
"query_pos_video_id": "v_yGCklVOLgk8",
"source": "pairs_with_reasoning",
"video_id": "v_yGCklVOLgk8"
}
|
1714394113.716848_fig
|
Amidst choppy waters, the man on the board touches up his hair.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst choppy waters, the man on the board touches up his hair.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man on a board in choppy water, with visible waves and splashes. At one point, the man adjusts his hair while maintaining balance on the board. Key elements from the query—choppy waters, the man on the board, and the action of touching up his hair—are all present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_3svMy4Kfjd8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.716848_fig::v_3svMy4Kfjd8",
"query_pos_video_id": "v_3svMy4Kfjd8",
"source": "pairs_with_reasoning",
"video_id": "v_3svMy4Kfjd8"
}
|
1714394113.693701_fig
|
A man, in board shorts, expertly maneuvers a wakeboard, following a boat's path on a body of water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man, in board shorts, expertly maneuvers a wakeboard, following a boat's path on a body of water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a man in board shorts (black with patterns) wearing a life vest and blue shirt, riding a wakeboard on the water’s surface created by a boat. The scene captures him maneuvering through the boat’s wake. Key elements from the query—“a man, in board shorts, expertly maneuvers a wakeboard, following a boat’s path on a body of water”—are all present: the man wears board shorts, uses a wakeboard, and moves along the wake generated by a boat.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_3svMy4Kfjd8.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.693701_fig::v_3svMy4Kfjd8",
"query_pos_video_id": "v_3svMy4Kfjd8",
"source": "pairs_with_reasoning",
"video_id": "v_3svMy4Kfjd8"
}
|
1714394113.718695_fig
|
Amidst the spray of water, a water skier in a life jacket maintains his balance on blue and white skis, being towed by a powerful white boat.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Amidst the spray of water, a water skier in a life jacket maintains his balance on blue and white skis, being towed by a powerful white boat.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows water spray around blue skis with white text and designs, a person holding a tow rope connected to a white boat creating a wake, and the skier maintaining balance. The query’s key elements—spray of water, blue and white skis, a white boat towing, and the skier’s balance—are all present. The skis match the blue-and-white description, the boat is white and powerful (evidenced by the wake), and the skier’s grip on the rope confirms being towed while balancing amid water spray.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_S7MZle1gPzI.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.718695_fig::v_S7MZle1gPzI",
"query_pos_video_id": "v_S7MZle1gPzI",
"source": "pairs_with_reasoning",
"video_id": "v_S7MZle1gPzI"
}
|
1714394113.730044_fig
|
A man skis behind a boat, connected by a rope and holding the handle.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man skis behind a boat, connected by a rope and holding the handle.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The video shows a first - person perspective of hands gripping a handle connected by a rope to a boat. Blue skis labeled “HQ SKIS” are visible in the water behind the hands. A boat with people is seen ahead, creating a wake. Key elements from the query (man skiing, rope connection, holding handle, boat) are all present in the visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_S7MZle1gPzI.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.730044_fig::v_S7MZle1gPzI",
"query_pos_video_id": "v_S7MZle1gPzI",
"source": "pairs_with_reasoning",
"video_id": "v_S7MZle1gPzI"
}
|
1714394113.684935_fig
|
A wakeboarder fine-tunes his bindings before being towed through the water.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A wakeboarder fine-tunes his bindings before being towed through the water.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a wakeboarder adjusting bindings (seen in the early frames where the individual is bent over the board), followed by being towed through water (evidenced by the rope, movement across the water, and splashes). Key elements from the query—binding adjustment and being towed through water—are all present in the video’s visual content.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_Tywf5EzWntY.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.684935_fig::v_Tywf5EzWntY",
"query_pos_video_id": "v_Tywf5EzWntY",
"source": "pairs_with_reasoning",
"video_id": "v_Tywf5EzWntY"
}
|
1714394113.682496_fig
|
Two individuals, one closer to the camera and the other further away, are water skiing while being pulled by a boat in a scenic lake setting.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"Two individuals, one closer to the camera and the other further away, are water skiing while being pulled by a boat in a scenic lake setting.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows two individuals water skiing on a lake, with one skier appearing closer to the camera and the other further away. The scene is set in a scenic lake environment with trees and boats visible in the background, matching the query’s description of the setting and the presence of two skiers at varying distances. Key elements (two skiers, differing proximity, lake setting with boats/trees) from the query are all present in the retrieved video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_Tywf5EzWntY.mkv"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.682496_fig::v_Tywf5EzWntY",
"query_pos_video_id": "v_Tywf5EzWntY",
"source": "pairs_with_reasoning",
"video_id": "v_Tywf5EzWntY"
}
|
1714394113.711149_fig
|
A man glides horizontally across the river, creating water spray, amidst choppy waters.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A man glides horizontally across the river, creating water spray, amidst choppy waters.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video shows a man water skiing across a river, which involves horizontal movement across the water. The skis generate visible water spray as he moves, and the river’s surface appears choppy with waves and turbulence. Key elements from the query—horizontal gliding, water spray, and choppy waters—are all present in the video.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_LdAT5MEBkuQ.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.711149_fig::v_LdAT5MEBkuQ",
"query_pos_video_id": "v_LdAT5MEBkuQ",
"source": "pairs_with_reasoning",
"video_id": "v_LdAT5MEBkuQ"
}
|
1714394113.695516_fig
|
A stationary man in sunglasses and a polo shirt, with a water ski, addresses the camera by a lake.
|
[
{
"content": "You are a multi-turn video retrieval assistant.\nYour goal is to determine whether a retrieved video matches the original query.\n\nYou will generate a latent query representation for retrieval.\nThis representation is not human-readable.\n\nOutput rules (MUST follow exactly):\n- Turn 1: output `New query:` followed by <SOFT_BOS> and exactly 8 soft tokens, then <SOFT_EOS>.\n- Turn 2+: always output <answer>matched</answer> or <answer>not_matched</answer> first.\n- If not matched, then output `New query:` with <SOFT_ADD> and exactly 4 soft tokens.\n- Always include a `Reasoning:` line explaining the decision or rewrite.\n- Do not invent details beyond the query or retrieved video content.\n",
"role": "system"
},
{
"content": "Instruction: Generate a retrieval-oriented query representation.",
"role": "user"
},
{
"content": "New query: <SOFT_BOS> <SOFT_0> <SOFT_1> <SOFT_2> <SOFT_3> <SOFT_4> <SOFT_5> <SOFT_6> <SOFT_7> <SOFT_EOS>\n",
"role": "assistant"
},
{
"content": "Retrieved video: <video>\nQuery: \"A stationary man in sunglasses and a polo shirt, with a water ski, addresses the camera by a lake.\"\nInstruction: Decide whether the video matches the query.",
"role": "user"
},
{
"content": "The retrieved video contains a stationary man in a white polo shirt and sunglasses holding a water ski (labeled “princesclub.com”) while speaking to the camera, with a lake, trees, and docks in the background. Key elements from the query—stationary man, sunglasses, polo shirt, water ski, lake setting, and addressing the camera—are all visually present in the retrieved scene.<answer>matched</answer>\n",
"role": "assistant"
}
] |
[
"/hub_data2/seohyun/softqmr/structured/toy_stage2/videos/v_LdAT5MEBkuQ.mp4"
] |
{
"episode_type": "T1",
"gold_label": "pos",
"pair_id": "1714394113.695516_fig::v_LdAT5MEBkuQ",
"query_pos_video_id": "v_LdAT5MEBkuQ",
"source": "pairs_with_reasoning",
"video_id": "v_LdAT5MEBkuQ"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.