RackLife: ten days in a GPU data center

RackLife is a long-video understanding benchmark: 50 hours of continuous footage that follows one technician, Alex, through ten working days at a fictional GPU data center (Helios Compute), with 592 questions about what happened.

The world is rendered procedurally in Panda3D from a scripted timeline, so every answer comes from ground truth rather than from human annotation of the footage. Every question was also checked against the rendered frames, so the evidence is visible on screen.

Video

• videos/day01.mp4 … videos/day10.mp4: one file per day, 10:00:00 to 15:00:00 game time, 5 hours each.
• 1280×720, 5 fps, H.264, 90,000 frames per day.
• Video second 0 is 10:00:00 of that day.
• The corner of every frame shows the day, the weekday and the clock. Day 1 is a Monday and Day 10 is a Wednesday.
• Spoken lines appear as subtitles at the bottom of the frame, prefixed with the speaker's name. Radio, phone, PA and lines heard through a door are marked. There is no audio track.
• The camera mostly follows Alex from behind. It cuts to close-ups when Alex looks closely at something, puts down or picks up an object, or watches CCTV playback.

Questions

Every question has a query time. A model may only see the video before the query time: all earlier days, plus the query day up to that moment. All evidence ends before the query time, and the answer never depends on anything that happens after it.

File	Format
questions_mc.json	Multiple choice, in the EgoLifeQA style. Fields: ID, query_time (date = DAYn, time = HHMMSSFF), type, src, question, choice_a … choice_g (only as many as the question has), answer (letter), evidence (per-day intervals in seconds from 10:00:00).
questions_open.json	Open answer, in the MM-Lifelong style. Fields: index, question, answer, question_type, src, query_time (video_id, seconds from 10:00:00, clock), clue_intervals.

The two files contain the same 592 questions.

No searchable on-screen text

The video is full of text: subtitles, door signs, rack and object labels, brand names on cans, screen contents and the clock. A caption index copies all of it word for word. If a question reused that text, a model could answer it with a keyword search, without understanding the video. So no question or option does:

• Spoken lines are paraphrased, or the question asks about what they imply, without their distinctive words. People's names are kept, because the subtitles show them.
• Rooms are named by how they look, for example "the dark gray room with a wall of camera monitors", not by their door signs. The four data halls look alike, so they are one kind of room: "a server hall with rows of black server racks".
• Objects are named by what they are and what happened to them, not by their labels. Soda is named by the color of the can.
• Times are given as parts of the day, such as "late in the morning", not as clock times.

Set 1: 339 questions

These are multi-hop retrieval questions anchored on something seen or heard. Examples:

• where I was when someone said something
• which room I walked into next
• what the red warning light was doing the next time it was in view
• what color of soda can I bought on the day someone handed me a part
• in what order three things happened

src is empty for these questions. All of them have four options.

Set 2: 253 questions

These follow the question types of four existing long-video benchmarks. src names the benchmark and type.

src	Count
EgoMemReason/Event Linking	65
MM-Lifelong/Counting	56
EgoMemReason/Temporal Counting	27
EgoLifeQA/EventRecall	23
EgoMemReason/Event Ordering	16
MM-Lifelong/Hallucination Detection	13
EgoLifeQA/RelationMap	11
EgoLifeQA/EntityLog	10
MM-Lifelong/Entity Recognition	8
EgoMemReason/Spatial Preference	7
EgoMemReason/Activity Pattern	7
MM-Lifelong/Attribute Recognition	5
MM-Lifelong/Temporal Reasoning	3
MM-Lifelong/Social Interaction	1
EgoLifeQA/HabitInsight	1

EgoMemReason-style questions have five to seven options, like that benchmark. All other questions have four.

How the questions were made and checked

1. Template generators compute each answer from the compiled timeline. They keep a question only when the per-second visibility log of the render shows that the evidence is on screen: the object, the person, the close-up or the subtitle.
2. A rewrite pass turned spoken lines into paraphrases. It also rewrote the hand-written questions under the rules above. A hand-written question that cannot be asked without on-screen text, for example one about a number shown on a screen, was dropped.
3. A program check confirms the logic of every question, and that its evidence ends before the query time. It also checks the final text. No sign name, label, brand or clock time may appear. No question may share four words in a row, two of them rare, with any subtitle or screen. No question may use a rare word from the spoken line it is built on.
4. A multi-agent review then checked every question against the frames, the event logs and the visibility logs, followed by an adversarial second pass. A question was dropped if it failed any of these criteria:
   • The evidence is not visible.
   • The anchor is ambiguous before the query time.
   • The answer is wrong.
   • A wrong option cannot be ruled out from the video.
   • The question cannot be answered from the video before the query time.
   • There is a shortcut: text shown earlier that gives the answer away, common sense, or an option that stands out by its format.
   • The question can be located or answered by searching on-screen text.

Of 1,078 candidates, 592 passed.