83K stratified subset: annotations, manifest, shard index, reader

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+annotations/all_sampled.json filter=lfs diff=lfs merge=lfs -text
+annotations/by_source/0_30_s_youtube_oe_v0_1_qa_processed.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,117 @@
+---
+license: other
+task_categories:
+  - video-text-to-text
+  - visual-question-answering
+language:
+  - en
+size_categories:
+  - 10K<n<100K
+tags:
+  - video
+  - long-video-understanding
+  - subset
+---
+
+# LLaVA-Video-83K
+
+An 83,000-sample subset of [`lmms-lab/LLaVA-Video-178K`](https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K),
+selected by stratified density-flattened sampling rather than uniformly at random,
+and shipped **with its media** so it is usable straight out of the bucket.
+
+| | |
+|---|---|
+| Records | **83,000** (5.0% of the 1,650,547 upstream records) |
+| QA pairs | 279,679 |
+| Distinct media | 71,224 (59,428 mp4 + 11,796 llava_hound frame scenes) |
+| Media size | 428.9 GiB |
+| Tasks | Caption 10.97% · Open-ended VQA 75.84% · Multiple-choice VQA 13.19% |
+
+Media for the `llava_hound` split comes from
+[`ShareGPTVideo/train_video_and_instruction`](https://huggingface.co/datasets/ShareGPTVideo/train_video_and_instruction)
+(`train_300k`), which covers all 179,318 referenced scenes.
+
+## Layout
+
+```
+LLaVA-Video-83K/
+├── annotations/
+│   ├── all_sampled.json          # 83,000 records, upstream schema unchanged
+│   └── by_source/*.json          # 39 files, split as upstream
+├── metadata/
+│   ├── manifest.parquet          # record -> media, task, duration bucket, source
+│   ├── shard_index.parquet       # member -> (shard, offset, nbytes)   ← random access
+│   ├── scene_index.parquet       # llava_hound scene -> frame count
+│   └── SAMPLING_REPORT.md        # how the subset was drawn, and its verification
+├── embeddings/                   # MiniLM-L6 vector per record, for re-drawing subsets
+├── shards/
+│   ├── video/video-NNNNN.tar     # uncompressed, ~2 GiB each
+│   └── frame/frame-NNNNN.tar     # uncompressed, ~2 GiB each
+└── videos_large/                 # the 188 mp4s above 64 MiB, stored loose
+```
+
+A record's `video` field is unchanged from upstream and is exactly the member
+name inside the shards, so nothing in the annotations needs rewriting.
+
+## Why tar shards
+
+Media is stored as **uncompressed** tar plus a byte-offset index, not as loose
+files. A member therefore occupies a contiguous byte range and is one `pread`
+away — `decord` decodes it from `BytesIO` byte-identically to reading the file
+from disk, seeks included (verified on files up to 707 MiB).
+
+The reason is I/O, not a Hub limit: 377k loose objects make transfer, listing
+and Lustre metadata operations painfully slow, while ~220 shards of 2 GiB read
+at full bandwidth. Videos above 64 MiB are the exception — they stay loose so a
+dataloader never has to buffer a 700 MiB file in RAM.
+
+## Usage
+
+```python
+from lvdu_reader import ShardReader
+
+r = ShardReader("/path/to/LLaVA-Video-83K")
+
+vr = r.video_reader("academic_source/Charades/RW587.mp4")   # decord VideoReader
+frames = r.scene_frames("shareVideoGPTV/frames/all_frames/v_XXX-Scene-001")
+```
+
+`ShardReader` is safe under DataLoader workers: shard handles are opened lazily
+per process and never shared across a fork.
+
+## How the subset was drawn
+
+Uniform random sampling inherits the upstream corpus's pathologies. This subset
+instead:
+
+- keeps the **task / duration / source mix** of the parent pool (task shares match to 0.01%);
+- gives small sources (activitynetqa, nextqa, perceptiontest — 41k records between them) a floor so they survive the cut;
+- drops records whose answers are entirely boilerplate, cutting ultra-short open-ended answers from 5.65% (random) to **2.39%**;
+- **flattens density** inside each stratum: cluster the MiniLM vectors, then allocate per cluster ∝ size^0.5, so dense topic clusters stop dominating;
+- constrains the multiple-choice **length shortcut**: a blind "pick the longest option" guesser scores 51.23% on the full set, and **49.79%** here.
+
+Measured effect of the flattening, against an equally-sized random draw: cluster
+*hit rate* is unchanged (the metric saturates at 5% sampling), but concentration
+drops — max samples per cluster 287 → **161** globally and 52 → **45** inside the
+largest stratum; Gini 0.712 → 0.697 and 0.394 → **0.341**.
+
+Full numbers, including the verification of every claim above, are in
+`metadata/SAMPLING_REPORT.md`.
+
+## Provenance and licence
+
+This is a derived subset. The videos and annotations remain under the terms of
+the upstream releases — see `lmms-lab/LLaVA-Video-178K` and
+`ShareGPTVideo/train_video_and_instruction`. Cite those datasets, not this
+repackaging, for the underlying data.
+
+## Media
+
+This repo carries the annotations and indices only. The 429 GiB of media
+lives in the companion bucket:
+
+```
+hf://buckets/LLMasterLL/LVDU/LLaVA-Video-83K/
+```
+
+Download it with `hf buckets sync`, then point `ShardReader` at the result.

annotations/all_sampled.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aae0844640379fa9c08dc53b97fc4af57b5500833c4ffaa0192bd689807d82f2
+size 105543456

annotations/by_source/0_30_s_youtube_oe_v0_1_qa_processed.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19b2e32979aec76b22ea7ea6c6b930c12e549af6192e86aa46f27d4fbab496bf
+size 22374179

annotations/by_source/2_3_m_nextqa_mc_qa_processed.json

@@ -0,0 +1 @@
+[{"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhere do the people carry out the activity?\nOptions:\nA. room.\nB. bar.\nC. ring.\nD. clinic.\nE. ice field.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "E."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nWhy did the man beside the man in green hit his hands in a uniform beat when the man is reciting?\nOptions:\nA. move out of the way.\nB. encourage.\nC. background for filming.\nD. pace the reciting.\nE. to dance.\n"}, {"from": "gpt", "value": "D. pace the reciting."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhat did the adult do to the dog when it approaches him?\nOptions:\nA. wag tail.\nB. run away.\nC. caress the dog.\nD. biting on to the blue toy.\nE. grab the mouth.\n"}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1013-10566292134", "conversations": [{"from": "human", "value": "<image>\nWhat do the three other people do as the man with glasses is speaking?\nOptions:\nA. take photos.\nB. look at him and listen.\nC. stand and admire the view.\nD. walks away.\nE. talk.\n"}, {"from": "gpt", "value": "B. look at him and listen."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/10566292134.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nHow did the man make sure that the baby is within his sight when moving around the house?\nOptions:\nA. push the baby walker.\nB. watch her crawl.\nC. put hand behind.\nD. not very stable.\nE. hands ready to catch the boy.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "B."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nHow did the boys feel when playing with the snow outdoors?\nOptions:\nA. happy.\nB. happy whenever the man kisses them.\nC. happy wagging tail.\nD. bored and sleepy.\nE. having fun.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "A. happy."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1016-11049178966", "conversations": [{"from": "human", "value": "<image>\nHow is the sound from the man s guitar projected when he is playing?\nOptions:\nA. with his hands.\nB. nodding his head.\nC. stamp her foot.\nD. move feet with the beat.\nE. microphone.\n"}, {"from": "gpt", "value": "E."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1016/11049178966.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nHow many cats are there?\nOptions:\nA. two.\nB. five.\nC. one.\nD. four.\nE. three.\n"}, {"from": "gpt", "value": "A."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nHow does the bigger dog shoo the smaller dog away in the middle of the video?\nOptions:\nA. on the floor.\nB. get up and turn around.\nC. caress dog.\nD. bites him.\nE. yawn.\n"}, {"from": "gpt", "value": "B. get up and turn around."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1013-10566292134", "conversations": [{"from": "human", "value": "<image>\nWhat does the man speaking do after first touching the white cup?\nOptions:\nA. pass paper to woman.\nB. keep his left arm on table.\nC. raise cup.\nD. walk forward and observe.\nE. took a step back.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "B. keep his left arm on table."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/10566292134.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhy does the helicopter fly for sometime with the person in red in the middle of the video?\nOptions:\nA. looking down at elephant.\nB. fly away.\nC. reach the correct destination.\nD. duck from attack.\nE. ball rolling away.\n"}, {"from": "gpt", "value": "C. reach the correct destination."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhat are the three boys looking at in the middle of the video?\nOptions:\nA. go back to the side.\nB. push bottles in front.\nC. their two other friends.\nD. turn behind to look.\nE. puts it in the box.\n"}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nHow many boys are there in the video?\nOptions:\nA. five.\nB. four.\nC. three.\nD. one.\nE. two.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "A. five."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhy does the boy wearing blue lie down on the ground?\nOptions:\nA. he wanted to escape.\nB. play his toy.\nC. make snow angel.\nD. agreeing with himself.\nE. ready to catch the toy.\n"}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhy are they performing?\nOptions:\nA. celebration.\nB. swaying while performing.\nC. music performance.\nD. enjoying music.\nE. engrossed in music.\n"}, {"from": "gpt", "value": "A."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nWhat does the smaller dog do after being pushed away by the bigger dog in the middle?\nOptions:\nA. watch.\nB. bite dog.\nC. takes the ball.\nD. keep turning around.\nE. runs toward the tutrle.\n"}, {"from": "gpt", "value": "D."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nHow does the bigger dog express its excitement while playing throughout the video?\nOptions:\nA. wag its tail.\nB. laugh.\nC. chase smaller dog.\nD. push the chair away.\nE. kick legs.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "C. chase smaller dog."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhy does the man in green point his hand to the men to his left in between the video?\nOptions:\nA. let him speak.\nB. look at the boy in stripes.\nC. playing with the tablet.\nD. communicating with boy in blue.\nE. see scenery closer via binocular.\n"}, {"from": "gpt", "value": "A. let him speak."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nWhat does the boy do after standing next to the girl near the end?\nOptions:\nA. touch her cheek with palm.\nB. stretch arms out.\nC. hold his hand.\nD. start dancing too.\nE. throw up the scarf.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "D."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nWhy did one of the cat walk over to the other box near the middle?\nOptions:\nA. go near camera.\nB. switch attention.\nC. play with spoon.\nD. enjoyed leaning on pillow.\nE. wants to enter.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "B. switch attention."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nWhy does the lady in blue hold up her phone?\nOptions:\nA. stylish.\nB. not get lost.\nC. pass it to boy later.\nD. listening to her.\nE. record for memory.\n"}, {"from": "gpt", "value": "E."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nHow does the cat check what is inside the smaller box?\nOptions:\nA. use mouth.\nB. use paw.\nC. stretches.\nD. jump up.\nE. lie on the carpet.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "B. use paw."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nWhy is the girl kneeling on the floor?\nOptions:\nA. feed ostrich.\nB. learn dancing.\nC. watching the dog and cat play.\nD. play with dog.\nE. push herself down.\n"}, {"from": "gpt", "value": "B. learn dancing."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nHow does the boy react after the girl lies facing down on the floor?\nOptions:\nA. tickle him.\nB. turn around.\nC. look at her.\nD. smiling.\nE. talk to the lady.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nHow did the boy protect himself from the cold outdoors?\nOptions:\nA. wear gloves.\nB. repeat sliding.\nC. wear winter wear.\nD. wear thick clothing.\nE. safety vest.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "D."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhy does the person hold the rope while coming down?\nOptions:\nA. acting as an animal.\nB. to hang the bottle.\nC. he slipped.\nD. to make man come down.\nE. lowered down slowly.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "E. lowered down slowly."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nHow does the man in red get down from the helicopter?\nOptions:\nA. throw with right hand.\nB. lowered down by rope.\nC. raise both his hands.\nD. controller.\nE. swing it around.\n"}, {"from": "gpt", "value": "B. lowered down by rope."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nHow were both cats using their paws at the start?\nOptions:\nA. touch each other.\nB. scratch against the box.\nC. climbs over it.\nD. by licking each other.\nE. licks the cat.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "B."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "1013-10566292134", "conversations": [{"from": "human", "value": "<image>\nHow does the man with glasses use body language to express himself when speaking?\nOptions:\nA. standing.\nB. use microphone.\nC. pressing buttons.\nD. hand gestures.\nE. upright.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "D."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/10566292134.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhy does the person in red raise his hand after tending to the person?\nOptions:\nA. tears.\nB. tourist.\nC. be pulled up.\nD. keep the beat.\nE. holding a stick.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "C. be pulled up."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhy did the boy in the middle pace around the stage before the performance?\nOptions:\nA. waiting his team member.\nB. wait for his turn.\nC. scared baby will fall off.\nD. prepare a performance.\nE. wants to jump again.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "A. waiting his team member."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nWhy did the man in green put down the piece of paper after the recites the entire script?\nOptions:\nA. record for later use.\nB. read to the audience.\nC. to catch his breath.\nD. finish reading.\nE. reading to girl.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "D."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nWhy are the cats jumping and surrounding the red box?\nOptions:\nA. just woke up.\nB. looking for food.\nC. curious.\nD. wants to go into box.\nE. play with it.\n"}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhat did the baby do when he approached a stack of toys near the start?\nOptions:\nA. pick up one and put into mouth.\nB. hit the wood.\nC. look at the doll.\nD. pull the toy in front.\nE. curious.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "A. pick up one and put into mouth."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nWhat does the bigger dog do after the smaller dog goes for its behind at the beginning?\nOptions:\nA. move towards big dog.\nB. lie down.\nC. caress the dog.\nD. chase after small dog.\nE. sniff it.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "B."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhat happens to the two people on the ground after the man in red raises up his hand?\nOptions:\nA. guiding.\nB. chewing something.\nC. get raised up by helicopter.\nD. put pieces on the board.\nE. hiking.\n"}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nHow does the child in white know the moves of the dance?\nOptions:\nA. walk past man.\nB. follow something.\nC. rubs the dog.\nD. remember in davance.\nE. clapping.\n"}, {"from": "gpt", "value": "B. follow something."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nHow many pieces of snow chunks did the boy in blue pick up?\nOptions:\nA. five.\nB. three.\nC. twice.\nD. four.\nE. seventeen.\n"}, {"from": "gpt", "value": "A."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nHow does the boy in green interact with the child in white near the end of the video?\nOptions:\nA. stretch his hand to point.\nB. pushed by boy in red.\nC. talking.\nD. rip off wrapping paper.\nE. dance together.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "E. dance together."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1013-10566292134", "conversations": [{"from": "human", "value": "<image>\nWhat does the man with glasses do after he first removed his hands from the cup?\nOptions:\nA. drink from the lid.\nB. does head bang.\nC. put cup on his face.\nD. put down the can.\nE. put his hands together.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "E."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/10566292134.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhere is the video taken?\nOptions:\nA. playground.\nB. classroom.\nC. outdoor.\nD. roadside.\nE. mountain.\n"}, {"from": "gpt", "value": "B. classroom."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nWhy did the small dog start running back and fourth in the middle of the video?\nOptions:\nA. play with each other.\nB. let the dog chase.\nC. sniffing the snow.\nD. play somewhere else.\nE. play catch.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "E. play catch."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhy are the boys walking around the tree?\nOptions:\nA. looking at the structure around.\nB. play game.\nC. only one place to hit ball.\nD. chicken went away from him.\nE. they are performing.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "E."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhat did the man in black do before he spoke?\nOptions:\nA. hold on rod.\nB. stretch his hands.\nC. take one leg off the scooter.\nD. smiling.\nE. look at man in white.\n"}, {"from": "gpt", "value": "D. smiling."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhy did the man in green stood up near the end of the video?\nOptions:\nA. to walk away.\nB. to bow together.\nC. finished shooting.\nD. to grill food.\nE. take pictures.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "A. to walk away."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhy does the person in blue place his finger under the tap?\nOptions:\nA. playing with the item.\nB. drink water.\nC. make cat comfortable with water.\nD. paint.\nE. cross the objects.\nPlease respond with only the letter of the correct answer."}, {"from": "gpt", "value": "C."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhy does the boy wearing blue carry big chunks of snow in his hand?\nOptions:\nA. play with the dog.\nB. play with snow.\nC. thick snow.\nD. look at the cameraman.\nE. play sledding.\n"}, {"from": "gpt", "value": "B. play with snow."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nWhy are the cats putting their paws inside the hole in the box?\nOptions:\nA. someone put her there.\nB. dog attack it.\nC. relax.\nD. curious what inside.\nE. getting ready to shower.\n"}, {"from": "gpt", "value": "D. curious what inside."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhat does the man in red do after reaching the ground?\nOptions:\nA. get raised up by helicopter.\nB. jumps.\nC. attach man on ground to him.\nD. throw something.\nE. raise his arm.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "C. attach man on ground to him."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nHow did the boy in red try to move his body when he is lying flat on the ground?\nOptions:\nA. act to kick him.\nB. lie on the ground.\nC. touched the baby s back.\nD. get up and dance.\nE. try to push himself forward.\nPlease provide your answer by stating the letter followed by the full option."}, {"from": "gpt", "value": "E. try to push himself forward."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhat did the man in purple do after the man in green raised his hand toward his left?\nOptions:\nA. smiles and waves.\nB. stand there.\nC. preparing to shoot.\nD. starts speaking.\nE. play his instrument.\n"}, {"from": "gpt", "value": "D."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}]

annotations/by_source/2_3_m_nextqa_oe_qa_processed.json

@@ -0,0 +1 @@
+[{"id": "1016-11049178966", "conversations": [{"from": "human", "value": "<image>\nWhat does the man do after putting the cup of water down?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "start playing the instruments."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1016/11049178966.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nWhat does the girl do after standing up?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "wave and dance."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhy does the boy wearing blue hit the ground with his hand and leg?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "play with snow."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1016-11049178966", "conversations": [{"from": "human", "value": "<image>\nWhere did the man put the cup after drinking?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "the case on his left."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1016/11049178966.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nWhy did the small dog crawl on the big dog from the start of the video to the end?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "play."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhat happens to the helicopter as it pulls up the two people?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "gets lower."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhy is the person coming down from the helicopter?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "save the person on the ice."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhy did the boy in red throw the snow chunk on the nearby tree after he stands up?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "want to see it break."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1016-11049178966", "conversations": [{"from": "human", "value": "<image>\nHow are the two men positioned on stage?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "standing."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1016/11049178966.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhere did the baby go to after touching the dog at first?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "same direction as dog."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nWhy does the girl wave her hands in the air after standing up?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "learn dancing."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhy is the lady in red observing the boys as they perform?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "she is the teacher."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nWhat did the man in green do after he finishes reading the script?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "places it down."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhy did the man point to the floor at the start of the video?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "tell baby to go there."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nHow did the man in white react after the man in green starts reading?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "turns to look at him."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhy did the baby mimick and caress the dog after the man pats it?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "copy whatever he sees."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhy does the cat sniff the green cup?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "searching for water."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nHow does the cat feel towards the running water?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "does not like it."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhy is the man in black moving his hands as he spoke near the end of the video?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "itchy."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhy did the man carry the baby and change her direction when the baby is crawling to the kitchen?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "dangerous."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhat did the first boy did when his two friends appear?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "sit in front of the tree."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nHow does the man in red protect himself?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "wear safety equipments."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nHow does the person turn on the water?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "using his hands."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1007-9020947679", "conversations": [{"from": "human", "value": "<image>\nWhat is the relationship of the 5 boys?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "classmates."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1007/9020947679.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nHow did the man in green amplified his voice?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "project his voice."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nHow did the boy in blue react when he saw the boy in red lie down and play on the snow?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "got up and hit the ground."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhy is the boy lying down on the pile of snow?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "play with snow."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhy did the man in black touched the microphone in the middle of the video?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "adjust the microphone."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhy did the man in purple turned back at the beginning of the video?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "check what shows on screen."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nWhy did the smaller dog start to play with the white dog toy at the middle of the video?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "distracted."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhy did the man look at the baby when she crawls around the house?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "ensure she is safe."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nHow does the bigger dog show affection towards the smaller dog at the end?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "lick its mouth."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nWhy did the man in white look at the man in green as he recites the script?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "interested to listen."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhat catches the cat s attention?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "flow of water."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nWhere is the ping pong ball?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "smaller red box."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nHow did the man in white show that he is engrossed in what the man in green is reciting?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "constantly look at him."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhy is the cat on the sink?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "drink water."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhat are the boys doing with the snow chunks on the ground?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "playing with them."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1006-9755380174", "conversations": [{"from": "human", "value": "<image>\nWhy does the child in white fall clap her hands when kneeling on the floor?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "dancing pose."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1006/9755380174.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nHow did the man change the baby s directions?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "carry the baby away."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nWhy did the baby slip and fall when trying to crawl out of the kitchen?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "lose balance."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nWhy did one of the cat jump to the other cat near the start?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "change hole."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhy does the person in red get hold of another person after landing on the ground?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "take that person up."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1016-11049178966", "conversations": [{"from": "human", "value": "<image>\nHow did the man get the cup of water at the start?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "picks the glass with left hand."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1016/11049178966.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhat does the person do after the cat looks at him?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "zooms into the tap."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhy does the cat stare into the sink after the tap turns on?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "gets surprised by water flow."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1013-10566292134", "conversations": [{"from": "human", "value": "<image>\nWhat was the man on the extreme right doing as the man with glasses held the cup and spoke?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "writing something."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/10566292134.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nHow did the three men showed visual images to aid their talk?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "presentation at the back."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhy does the person in blue turn on the tap?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "let cat drink water."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1110-12398010263", "conversations": [{"from": "human", "value": "<image>\nHow is the big brown dog related to the small puppy?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "parent offspring."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1110/12398010263.mp4"}, {"id": "1021-10440265715", "conversations": [{"from": "human", "value": "<image>\nWhy did the boy in red start to throw the snow after picking up on his hands?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "play with snow."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1021/10440265715.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nHow did the man in green feel while reading the scripts?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "intrigued."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}, {"id": "1107-13517946454", "conversations": [{"from": "human", "value": "<image>\nWhat event is going on in the video?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "book launch."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1107/13517946454.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhy does the person in blue control the amount of water coming out of thetap?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "save water."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1013-13144870414", "conversations": [{"from": "human", "value": "<image>\nHow does the baby move around the house?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "crawl."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1013/13144870414.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nHow does the person interact with the running water near the end?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "runs her hand through it."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "1099-9610744135", "conversations": [{"from": "human", "value": "<image>\nWhy did the cat have its head in the smaller red box?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "see what inside."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1099/9610744135.mp4"}, {"id": "0084-10522689523", "conversations": [{"from": "human", "value": "<image>\nWhere is the cat hanging out?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "toilet."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0084/10522689523.mp4"}, {"id": "0075-12843838795", "conversations": [{"from": "human", "value": "<image>\nWhy does the person come down slowly?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "safety."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/0075/12843838795.mp4"}, {"id": "1106-9038264963", "conversations": [{"from": "human", "value": "<image>\nWhy did the man in green read the script on his hands out loud?\nAnswer the question using few words or phrase."}, {"from": "gpt", "value": "recite the script."}], "data_source": "2_3_m_nextqa", "video": "NextQA/NExTVideo/1106/9038264963.mp4"}]

lvdu_reader.py

@@ -0,0 +1,127 @@
+"""Read LLaVA-Video-83K media out of the tar shards.
+
+The dataset stores media as uncompressed tar shards plus a byte-offset index, so
+any file is one pread away and nothing has to be unpacked before training.
+
+    from lvdu_reader import ShardReader
+    r = ShardReader("/path/to/LLaVA-Video-83K")
+
+    # annotation "video" field goes straight in
+    vr = r.video_reader("academic_source/Charades/RW587.mp4")   # decord VideoReader
+    frames = r.scene_frames("shareVideoGPTV/frames/all_frames/v_XXX-Scene-001")
+
+Videos larger than 64 MiB are stored loose under videos_large/ and are handed to
+decord as a path, so a 707 MiB file is streamed from disk rather than buffered
+in RAM. Everything else comes back as bytes via BytesIO, which decord decodes
+byte-identically (verified against on-disk reads, seeks included).
+
+Safe under DataLoader workers: file handles are opened lazily per process and
+cached, and are never shared across a fork.
+"""
+from __future__ import annotations
+
+import io
+import os
+
+HOUND_PREFIX = "shareVideoGPTV/frames/all_frames/"
+
+
+class ShardReader:
+    def __init__(self, root: str):
+        import pandas as pd
+        self.root = root
+        idx = pd.read_parquet(os.path.join(root, "metadata", "shard_index.parquet"))
+        self._idx = {m: (s, o, n) for m, s, o, n in
+                     zip(idx.member, idx.shard, idx.offset, idx.nbytes)}
+        self._kind = dict(zip(idx.member, idx.kind))
+        # scene -> sorted frame members, built once. Scanning all ~377k keys per
+        # call would be O(n) inside the dataloader hot path.
+        self._scene: dict[str, list[str]] = {}
+        for m, k in zip(idx.member, idx.kind):
+            if k == "frame":
+                self._scene.setdefault(m.split("/", 1)[0], []).append(m)
+        for v in self._scene.values():
+            v.sort()
+        self._fh: dict[str, io.BufferedReader] = {}
+        self._pid = os.getpid()
+
+    # ---------------------------------------------------------------- internals
+
+    def _handle(self, shard: str, kind: str) -> io.BufferedReader:
+        if os.getpid() != self._pid:      # forked into a worker: drop inherited fds
+            self._fh.clear()
+            self._pid = os.getpid()
+        fh = self._fh.get(shard)
+        if fh is None:
+            sub = "video" if kind == "video" else "frame"
+            fh = open(os.path.join(self.root, "shards", sub, shard), "rb", buffering=0)
+            self._fh[shard] = fh
+        return fh
+
+    def _resolve(self, key: str) -> str:
+        """Annotation `video` field -> index member key."""
+        return key[len(HOUND_PREFIX):] if key.startswith(HOUND_PREFIX) else key
+
+    # ------------------------------------------------------------------- public
+
+    def __contains__(self, key: str) -> bool:
+        # A llava_hound `video` field names a *scene directory*, not a file, so
+        # it is never itself an index member -- its frames are "<scene>/x.jpeg".
+        k = self._resolve(key)
+        return k in self._idx or k in self._scene
+
+    def kind_of(self, key: str) -> str:
+        """'video' or 'scene' -- which accessor this key needs."""
+        k = self._resolve(key)
+        if k in self._scene:
+            return "scene"
+        if k in self._idx:
+            return self._kind[k]
+        raise KeyError(key)
+
+    def open(self, key: str):
+        """Whatever this key represents: a decord VideoReader or a frame list."""
+        return self.scene_frames(key) if self.kind_of(key) == "scene" \
+            else self.video_reader(key)
+
+    def path_of(self, key: str) -> str | None:
+        """Filesystem path if the member is stored loose, else None."""
+        member = self._resolve(key)
+        shard, _, _ = self._idx[member]
+        return os.path.join(self.root, "videos_large", member) if shard == "" else None
+
+    def read(self, key: str) -> bytes:
+        """Raw bytes of one member (one pread for packed members)."""
+        member = self._resolve(key)
+        shard, offset, nbytes = self._idx[member]
+        if shard == "":
+            with open(os.path.join(self.root, "videos_large", member), "rb") as fh:
+                return fh.read()
+        fh = self._handle(shard, self._kind[member])
+        return os.pread(fh.fileno(), nbytes, offset)
+
+    def video_reader(self, key: str, **kw):
+        """decord.VideoReader for a video member; loose files go by path."""
+        from decord import VideoReader, cpu
+        kw.setdefault("ctx", cpu(0))
+        p = self.path_of(key)
+        return VideoReader(p, **kw) if p else VideoReader(io.BytesIO(self.read(key)), **kw)
+
+    def scene_members(self, key: str) -> list[str]:
+        """Sorted frame members of a llava_hound scene."""
+        scene = self._resolve(key)
+        try:
+            return self._scene[scene]
+        except KeyError:
+            raise KeyError(f"unknown scene {scene}") from None
+
+    def scene_frames(self, key: str) -> list:
+        """Decoded PIL frames of a llava_hound scene."""
+        from PIL import Image
+        return [Image.open(io.BytesIO(self.read(m))).convert("RGB")
+                for m in self.scene_members(key)]
+
+    def close(self) -> None:
+        for fh in self._fh.values():
+            fh.close()
+        self._fh.clear()

metadata/SAMPLING_REPORT.md

@@ -0,0 +1,65 @@
+# 83K 子集 — 抽样结果与验收
+
+生成于 `05_select.py`（seed=0, alpha=0.5, floor=300）。
+
+## 硬检查
+
+- 记录数 **83,000**（目标 83,000）— ✅
+- 不同媒体 **71,224** 个，落盘 **428.9 GiB**
+- 缺失媒体 **0** 个，零字节 **0** 个 — ✅
+- QA 对 **279,679**
+
+## 分布对照
+
+![](out/figs/cmp_shares.png)
+
+| cohort   |   cap_pairs |   cap_median_wc |         oe_pairs |   oe_median_awc |   oe_short_le3 |   oe_dup_q |
+|:---------|------------:|----------------:|-----------------:|----------------:|---------------:|-----------:|
+| full     |      178510 |             308 |      4.66152e+06 |              15 |         0.0591 |     0.3337 |
+| selected |        9103 |             327 | 229634           |              15 |         0.0239 |     0.2172 |
+| random   |        9090 |             303 | 236820           |              15 |         0.0565 |     0.2245 |
+
+
+## MC 长度捷径（盲选最长选项的准确率）
+
+| 全量 | 同规模随机 | **本子集** |
+|---|---|---|
+| 51.23% | 51.01% | **49.79%** |
+
+守门阈值是全量 +1.0 个百分点；实际发生 0 次换入。
+
+## 向量空间覆盖度
+
+三个粒度下对比"覆盖式抽样"和"同规模纯随机"（`09_coverage_probe.py`）：
+
+| 粒度 | 子集 | 命中簇 | 命中率 | Gini | 单簇最大样本数 |
+|---|---|---|---|---|---|
+| 全池 k=2000 | 覆盖式 | 1,753 | 87.6% | **0.541** | — |
+| | 随机 | 1,748 | 87.4% | 0.561 | — |
+| 全池 k=8300 | 覆盖式 | 5,106 | 61.5% | **0.697** | **161** |
+| | 随机 | 5,037 | 60.7% | 0.712 | 287 |
+| 最大层 k=2004<br>(oe/0-30s/youtube, 418K→20,045) | 覆盖式 | 1,950 | 97.3% | **0.341** | **45** |
+| | 随机 | 1,941 | 96.9% | 0.394 | 52 |
+
+**结论要分开看，不能笼统说"覆盖更好"：**
+
+- **命中率上没有意义的差别**（61.5% vs 60.7%、97.3% vs 96.9%）。
+  在 5% 采样率下这个指标已经饱和 —— 随机抽也能碰到绝大多数簇，
+  所以"命中多少个簇"根本区分不出两种策略。一开始就不该拿它当主指标。
+- **抗过度集中上有真实收益**：单簇最大样本数从随机的 287 降到 161（全池）、
+  52 降到 45（层内）；Gini 从 0.712 → 0.697、0.394 → 0.341。
+  这才是 `alpha=0.5` 实际起作用的地方 —— 它压的是密集题材簇的过采样，
+  不是去开拓新区域。
+
+顺带记录一个被这个探针抓出来的 bug：最初的余数分配把 floor 之后的缺口
+一次性全给了"剩余空间最大"的那个簇，结果子集在单簇里堆了 **775** 个样本
+（随机只有 52），等于把展平反向做成了集中。改成最大余数法逐个发放后才有上表的结果。
+
+## 产物
+
+```
+annotations/   与上游同 schema，按原文件切分 + all_sampled.json
+videos/        mp4，保持原相对路径
+frames/        llava_hound 的帧目录
+manifest.parquet | .csv
+```

metadata/coverage_probe.json

@@ -0,0 +1,37 @@
+{
+  "global_k8300": {
+    "selected": {
+      "hit": 5106,
+      "of": 8300,
+      "hit_frac": 0.6151807228915662,
+      "gini": 0.6965150791116272,
+      "max_per_cluster": 161
+    },
+    "random": {
+      "hit": 5037,
+      "of": 8300,
+      "hit_frac": 0.6068674698795181,
+      "gini": 0.7120930091450138,
+      "max_per_cluster": 287
+    }
+  },
+  "stratum_oe_0_30_s_youtube": {
+    "stratum_size": 418226,
+    "picked": 20045,
+    "k": 2004,
+    "selected": {
+      "hit": 1950,
+      "of": 2004,
+      "hit_frac": 0.9730538922155688,
+      "gini": 0.34122543140209977,
+      "max_per_cluster": 45
+    },
+    "random": {
+      "hit": 1941,
+      "of": 2004,
+      "hit_frac": 0.968562874251497,
+      "gini": 0.3937621140856227,
+      "max_per_cluster": 52
+    }
+  }
+}

metadata/manifest.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5d2389df6870f70c309e809ad60c16e4ff91c0263072779359c400d8a413fbf
+size 3423550

metadata/scene_index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f09b82400a54dcbab3d155f67093835402d5c480bff11edacfc37df80bddb89b
+size 118242

metadata/selection_report.json

@@ -0,0 +1,33 @@
+{
+  "target": 83000,
+  "selected": 83000,
+  "alpha": 0.5,
+  "floor": 300,
+  "seed": 0,
+  "distinct_media": 71224,
+  "media_GiB": 428.883671566844,
+  "by_task": {
+    "cap": 9103,
+    "mc": 10949,
+    "oe": 62948
+  },
+  "by_bucket": {
+    "0_30_s": 30439,
+    "1_2_m": 13460,
+    "2_3_m": 12460,
+    "30_60_s": 14636,
+    "llava_hound": 12005
+  },
+  "by_family": {
+    "academic": 12917,
+    "activitynetqa": 1822,
+    "llava_hound": 12005,
+    "nextqa": 3161,
+    "perceptiontest": 646,
+    "youtube": 52449
+  },
+  "qa_pairs": 279679,
+  "mc_pick_longest": 0.497861202025615,
+  "mc_pick_longest_full": 0.5123086448198672,
+  "mc_swaps": 0
+}

metadata/shard_index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:488327c32f518680dc5ec22400220c9ec572e1bbbe70ed14aaff1ac33fcd4bd6
+size 3083935