README.md

metadata

license: cc-by-4.0
task_categories:
  - video-text-to-text
tags:
  - benchmark
  - multimodal retrieval
size_categories:
  - n<1K

🤗 About This Repo

This repository hosts a representative sample of the FLARE (FLARE: Full-Modality Long-Video Audiovisual Retrieval Benchmark with User-Simulated Queries) benchmark for OpenReview reviewers. The full benchmark contains 399 long-form videos (225.4 h, 10–60 min each) screened from Video-MME, segmented into 87,697 fine-grained clips and annotated with vision / audio / unified audiovisual captions, plus 274,933 user-simulated queries (86,350 vision-only, 135,003 audio-only, 53,580 cross-modal). The full release lives at YqjMartin/FLARE; this demo lets reviewers inspect data quality without downloading the ~71 GB full dataset.

---

🧭 How These Samples Were Selected

From the subset of clips that retain at least one unified query after the hard-bimodal-constraint filtering, we enumerated all triplets of three temporally consecutive such clips and randomly drew two, which became sample_0 and sample_1. We chose this slice because unified cross-modal queries are the smallest and most discriminating part of FLARE (53,580 vs. 86,350 vision-only and 135,003 audio-only queries), and three consecutive clips give a similar-but-distinct local context that better reflects the disambiguation difficulty FLARE is designed to probe than any single clip viewed in isolation. For each triplet we release the full source video, the three clip videos, and every JSONL artefact the FLARE pipeline produces for those clips and full video. Note: in both samples one clip (Scene-068 in sample_0, Scene-040 in sample_1) carries no vision-only query — its candidate vision queries failed the stage-2 rank-1 retrieval check against the vision gallery and were dropped, even though the corresponding unified query passed the hard bimodal constraint, illustrating in vivo that vision alone can be insufficient for clips whose distinguishing signal is audiovisual.

---

🍺 About a Concrete Case — Sample 0

Clip	Unified query (cross-modal, hard-bimodal)	Unified caption (clip-level)
👆 Scene-067	"A woman in a blue scarf points off-screen while another person joins her, and a news-style voice says something about a president's promise near Antakya"	Show full unified caption The video clip features two individuals standing outdoors in front of a construction site. The setting appears to be overcast, with muted lighting suggesting a cool or cloudy day. In the background, several multi-story buildings are visible — some are unfinished concrete structures with exposed window frames, while others are covered in scaffolding, indicating active construction. The ground in the foreground is uneven and muddy, with orange safety fencing partially visible, demarcating the construction zone. The two individuals are positioned side by side, facing slightly to the right of the frame. The person on the left is a man with short dark hair and a beard, wearing a black jacket with a blue hood visible underneath. The person on the right is a woman with shoulder-length blonde hair, wearing a dark coat and a bright blue scarf. She is actively gesturing with her left arm extended outward, pointing toward something off-screen to the right. Her expression is engaged, and she appears to be speaking or explaining something. As the clip progresses, the man turns his head slightly toward the woman and then extends his own right arm, pointing in the same direction she indicated. His gesture suggests he is either confirming her point or adding to her explanation. The woman lowers her arm slightly but continues to look in the direction they are both pointing. Their body language indicates a collaborative or conversational interaction, possibly a guided tour or site inspection. The camera remains relatively steady throughout the clip, maintaining a medium shot that captures both individuals from the waist up, with the construction site clearly visible behind them. There is no significant camera movement or change in framing, keeping the focus on the interaction between the two people and their gestures toward the background environment. The DW logo is visible in the top left corner, indicating the footage is from Deutsche Welle. The audio clip features a single female speaker delivering a narrative in a clear, formal tone, characteristic of a news report or documentary. She speaks in English with a neutral accent. The content of her speech is: "originally promised by the president. Here on the outskirts of Antakya, Mustafa…" The speech is abruptly cut off mid-sentence, with the audio ending suddenly while she is still speaking. Throughout the clip, there is a persistent, low-frequency hum in the background, suggesting the presence of machinery or an engine. Additionally, there is a faint, high-pitched, rhythmic clicking sound, possibly from a vehicle's turn signal or a similar mechanical source. These background sounds are consistent and do not change in intensity or character during the recording. There is no music present.
👆 Scene-068	"overhead shot of a muddy construction zone with a van and a gazebo, while a voice talks about a team led by Arslan"	Show full unified caption The video presents a wide-angle, elevated view of a residential construction site featuring several multi-story apartment buildings in various stages of completion. The buildings are modern in design, with flat facades, rectangular windows, and a color scheme that includes white, gray, orange, and beige panels. Some buildings have red-tiled roofs, while others appear to be under roofing or finishing work. Scaffolding is visible on one of the central buildings, indicating ongoing exterior work. In the foreground, there is an unpaved, muddy area with newly planted young trees supported by wooden stakes. A few small, conical evergreen trees are spaced evenly across the area. A paved walkway or driveway runs diagonally through the scene, with several vehicles parked on it — including a white sedan, a silver SUV, and a white van. A small wooden gazebo with a green roof is visible on the right side of the frame, partially obscured by a white building edge. Multiple construction workers wearing high-visibility vests (orange and yellow) are scattered throughout the site. Some are walking along the driveway, while others are engaged in tasks near the buildings or around construction materials. A blue tarp covers a pile of materials or equipment near the center-right, and an orange excavator arm is partially visible beneath it. Piles of dirt, construction debris, and wooden planks are strewn across the ground, indicating active landscaping and infrastructure work. The camera remains static throughout the clip, offering a continuous, observational perspective. There is no significant change in the positions of the buildings or major structures, but minor movements are visible among the workers. The overall scene conveys a sense of active development and transition, with the built environment in the process of being completed and landscaped. The sky is overcast, casting diffuse, even lighting across the scene. The DW logo is visible in the top-left corner, indicating the source of the footage. The audio contains a single speaker, a female with a clear, articulate voice and a neutral accent. She is speaking in English, delivering a narrative or informational statement. The content of her speech is: "Arslan and his team from the state housing agency TOKI." The recording is of high quality, with no discernible background noise, music, or sound effects. The speaker's voice is the sole focus, and the audio is clean and free of distortion or static. The delivery is steady and measured, suggesting a formal or documentary-style narration. There are no changes in the audio environment or speaker characteristics throughout the clip.
👆 Scene-069	"workers in vests moving on a half-built building while a female speaker talks about getting apartments ready quickly"	Show full unified caption The video shows a multi-story building under construction, enveloped in metal scaffolding that covers its entire exterior. The building appears to be several stories high, with visible window openings on each floor. The facade is partially finished: some sections are covered with white insulation panels studded with black fasteners, while others reveal orange or yellow insulation material beneath. The structure's concrete frame is visible in places where the cladding has not yet been applied. Several construction workers are visible on different levels of the scaffolding. One worker, wearing a vest and dark pants, is seen on a middle floor, moving slightly within a window opening. Another worker, dressed in dark clothing, is positioned on a lower level. A third worker, wearing a gray shirt and pants, is visible higher up on the left side of the building, standing on a scaffold platform. The scaffolding is a dense grid of metal poles and cross-bracing, with horizontal platforms at each floor level. Some platforms have tools, buckets, or construction materials placed on them. A few pieces of fabric or clothing are draped over parts of the scaffolding, possibly for drying or as makeshift coverings. In the background, beyond the construction site, a residential area is visible, including houses with red-tiled roofs and trees. The sky is overcast, casting a diffuse, even light over the scene. The ground around the building is uneven and covered with dirt, construction debris, and scattered materials. Throughout the video, the workers make small movements — adjusting positions, reaching for tools, or shifting their stance — but there is no dramatic or rapid motion. The camera remains mostly static, with only slight adjustments in framing, maintaining a medium-long shot that captures the full height of the building and part of the surrounding environment. The overall impression is one of ongoing, methodical construction work in a semi-urban setting. The audio features a single female speaker with a clear, articulate voice and a neutral accent. She is speaking in a formal, narrative tone, as if reading from a script or delivering a report. The speech content is: "are racing to get the first few thousand apartments ready." The recording is of high quality, with no discernible background noise, music, or ambient sounds. The environment appears to be a controlled studio setting. The speaker's voice is steady and consistent in volume and pace throughout the short clip. There are no other audible elements or changes in the audio composition.

Full source video: K759eXmaMTY.mp4

---

🍺 About a Concrete Case — Sample 1

Clip	Unified query (cross-modal, hard-bimodal)	Unified caption (clip-level)
👆 Scene-040	"rotating view of a realistic heart model with a perforated dome device while a narrator mentions bridged approval for special hearts"	Show full unified caption The video presents a close-up, computer-generated 3D model of a human heart with a mechanical device attached to its lower chamber. The heart is rendered in a realistic flesh tone, with visible major blood vessels extending from the top. The mechanical device is a smooth, white, dome-shaped apparatus with a pattern of small, evenly spaced circular perforations across its surface. It is securely connected to the heart via a metallic, cylindrical interface that appears to be bolted or clamped to the heart's exterior. The video begins with a tight focus on the junction between the heart and the device, highlighting the mechanical attachment point. As the video progresses, the camera slowly pulls back and rotates slightly around the object, revealing more of the heart's structure and the full curvature of the device. The heart remains stationary throughout, while the camera movement provides a dynamic perspective, allowing viewers to observe the device's placement and integration with the organ. The background is a plain, bright white, ensuring that all visual attention remains on the heart and the attached device. There are no other objects or characters present in the scene. The lighting is soft and even, casting subtle shadows that accentuate the contours of both the heart and the device. In the top left corner, the word "Carmat" is visible, and in the top right corner, a small circular logo appears — both remain static throughout the video. The overall impression is clinical and technical, emphasizing the design and integration of the artificial device with the biological organ. The video does not depict any motion of the heart or device, suggesting it is a static demonstration rather than a simulation of function. The audio features a single male speaker with a clear, mid-to-low pitched voice. He is speaking in a calm, narrative tone, delivering a continuous sentence. The speech is in English and appears to be part of a documentary or informational recording. The content of the speech is: "special hearts in the world right now. But they're currently approved only as bridged." Throughout the entire clip, there is a continuous, low-frequency electronic hum in the background. This hum is steady and consistent, with no discernible changes in pitch or volume. There are no other background sounds, such as ambient noise, music, or sound effects. The recording is clean, with no noticeable static, distortion, or clipping. The speaker's voice is prominent and clear, suggesting a close microphone placement or professional recording setup.
👆 Scene-041	"A patient in a dotted hospital gown being helped by two doctors in a hospital room, with someone talking about temporary hearts in the background."	Show full unified caption The video opens with a bright, overexposed frame that quickly transitions into a clear, well-lit hospital setting. The scene is set in what appears to be an intensive care unit (ICU), as indicated by the "ICU 78" sign visible on the wall in the background. The environment is clinical, with medical equipment, IV poles, and monitors present. The "SynCardia" logo is visible in the top left corner of the frame. Three main individuals are prominently featured in the foreground. On the left is a male healthcare worker wearing light blue scrubs, with a stethoscope around his neck and an ID badge clipped to his chest. He is focused on a large, portable medical monitor with a touchscreen interface, which he holds with both hands. His attention is directed downward at the screen, suggesting he is reviewing or adjusting patient data. In the center stands a male patient wearing a light green hospital gown with dark polka dots. He is bald, appears to be middle-aged, and has a calm but attentive expression. His right arm is wrapped in a white bandage or compression sleeve, and he is holding onto the monitor's handle with his left hand, indicating he may be standing with support or participating in a mobility or rehabilitation activity. On the right is a female healthcare worker in dark green scrubs, also wearing gloves and an ID badge. She is looking down, possibly at the patient's arm or at a device she is holding, and appears to be assisting or monitoring the patient's condition. Her posture is attentive and professional. In the background, other medical staff and patients are visible, including a woman in pink scrubs who briefly walks into the frame and looks toward the central group. The background also includes medical equipment such as IV bags hanging from poles, monitors, and a hospital bed partially visible on the right. The camera remains relatively steady, with slight movements to follow the actions of the central figures. The focus remains on the interaction between the patient and the two healthcare workers, emphasizing the clinical care being provided. The spatial arrangement suggests a collaborative effort, with the patient positioned between the two medical professionals, who are both engaged in monitoring or assisting him. Throughout the video, the lighting remains consistent, and the environment conveys a sense of routine medical activity. The audio contains a single male speaker with a clear, steady voice, delivering a spoken narration. He is discussing medical technology, specifically mentioning "transplant devices" and "temporary hearts" that are used to sustain patients. The narration is cut off mid-sentence. There is no background music or ambient noise. The recording is clean and focused solely on the speaker's voice.
👆 Scene-042	"Surgical team in blue gowns working intently while a man says the heart is ready, followed by a deep synth tone"	Show full unified caption The video captures a scene inside a brightly lit, modern operating room during what appears to be a complex surgical procedure. The environment is sterile and highly organized, with medical equipment and monitors visible throughout the space. At the center of the frame, a group of at least six medical professionals — likely surgeons and surgical assistants — are clustered closely around the patient, who is not directly visible but is implied to be on the operating table beneath them. All personnel are dressed in full surgical attire: light blue disposable gowns, caps, face masks, and gloves. Their backs are mostly to the camera, and they are leaning forward, intensely focused on the surgical site. Their movements are coordinated and deliberate, suggesting a high level of concentration and teamwork. Above the surgical team, two large, circular surgical lights hang from the ceiling, casting a bright, focused illumination on the operating area. To the right, a large monitor displays real-time vital signs and graphical data, including heart rate and other physiological metrics. Smaller monitors and medical devices are also visible along the walls and on mobile carts. In the foreground, a metal instrument tray is positioned on a table covered with a blue sterile drape. The tray contains an array of surgical tools — forceps, clamps, scissors, and other metallic instruments — neatly arranged and ready for use. Additional items such as plastic bowls, syringes, and gauze are also visible on the table. The camera remains mostly stationary throughout the clip, with only slight adjustments in angle or zoom, maintaining a consistent view of the surgical team and their immediate surroundings. The depth of the scene is emphasized by the layered arrangement of personnel, equipment, and background elements — the surgical team in the mid-ground, the instrument tray in the foreground, and monitors and wall-mounted equipment in the background. The audio clip begins with a single male speaker delivering a short, clear sentence. His voice is calm and steady, with a neutral, narrative tone. He says, "The real donor heart becomes available." Immediately following the spoken sentence, the audio transitions to a sustained, low-pitched musical chord. This chord is produced by a synthesizer or electronic organ and has a smooth, ambient quality. It continues without any change in pitch or volume until the end of the clip. There are no other discernible background sounds, music, or additional speakers.

Full source video: 3uV8XZcIBbk.mp4

---

Each sample_*/ directory additionally provides per-modality clip captions / queries (clip-caption-{vision,audio,unified}.jsonl, clip-query-{vision,audio,unified}.jsonl) and video-level captions (video-caption-{vision,audio,unified}.jsonl), all in the exact JSONL schema of the full FLARE release.