The Dataset Viewer has been disabled on this dataset.

MMSB sources: the media MultiMediaSearchBench's scenarios are built from

MultiMediaSearchBench (MMSB, once called MultiMediaMemoryBench, which is where this repository's name comes from) measures how well an AI agent finds things in a large pile of someone's media. Each scenario builds its pile, and its questions, from a real dataset's own files and labels. This repository holds what 16 of the 23 scenarios read, laid out as MMSB expects: the sources' own media and labels. MMSB's code writes the questions from them.

The licences differ from folder to folder, and many of them allow non-commercial use only. Check the table before you use a folder.

Folder Scenario What it is Size Licence
meva/ v1_security: three outdoor security cameras MEVA: cameras G336, G341 and G436 on 12 March 2018, 10:00-11:15 (5-minute AVI clips), with Kitware's two activity label sets 5.9 GB CC BY 4.0
st2-followcam/ v3_match: football match clips SoccerTrack v2 follow-cam event clips: 525 thirty-second clips of three matches, each with its labels 8.3 GB CC BY 4.0
ami/ v4_meetings: work meetings AMI Meeting Corpus: recorded design meetings with their room sound, cameras and annotations 3.7 GB CC BY 4.0
ocw-14-01/ v9_lectures: lecture recordings MIT OpenCourseWare 14.01, Fall 2023: 26 lecture videos and their captions 2.9 GB CC BY-NC-SA 4.0
caltech-camera-traps/ i3_trailcam: a trail camera's memory card Caltech Camera Traps: location 23's 4,256 photos, with the dataset's labels 2.0 GB CDLA-Permissive 1.0
bd3-building-defects/ i4_site_photos: building inspection photos BD3: 793 labelled photos of building defects 0.3 GB CC BY 4.0
sroie/ i7_receipts: receipts and bills ICDAR 2019 SROIE: scanned Malaysian shop receipts 0.5 GB CC BY 4.0
cord-v2/ i7_receipts: receipts and bills CORD v2: Indonesian cafe receipts 2.3 GB CC BY 4.0
ocw-pages/ i8_course_pages: course notes and slides The lecture notes and slides of four MIT OpenCourseWare courses, as PDFs 0.1 GB CC BY-NC-SA 4.0
inat-observer/ i9_nature: nature photos 3,240 iNaturalist photos by one observer, with one row per photo of what it shows and when. No places and no EXIF GPS 1.5 GB CC BY-NC, photo by photo
icsi/ a3_meetings: research group meetings ICSI Meeting Corpus: ten meetings' room recordings, with transcripts and annotations 1.2 GB CC BY 4.0
birdvox-full-night/ a4_field: a field recorder BirdVox-full-night: unit 3's night of audio, with the annotations of six units 0.9 GB CC BY 4.0
ocw-6-0002/ a6_lecture_audio: recorded lectures MIT OpenCourseWare 6.0002, Fall 2016: 15 lectures and their captions 1.7 GB CC BY-NC-SA 3.0
chalk-radio/ a7_podcast: a podcast backlog Chalk Radio, MIT OpenCourseWare's podcast: 60 episodes and their captions 3.2 GB CC BY-NC-SA 4.0
fma/ a8_music: a music library Free Music Archive: fma_small.zip (8,000 thirty-second clips) and fma_metadata.zip 8.0 GB each track its artist's (see below); metadata CC BY 4.0
robocall-ftc/ m3_voicemail: voicemails 1,427 robocall recordings from FTC cases, with transcripts and the case papers 1.2 GB CC BY-ND 4.0 taken (see below)
fsd50k/ m4_samples: a folder of sound clips FSD50K: the evaluation split's 10,231 clips, with its labels, vocabulary, ratings and per-clip licences 8.9 GB CC BY 4.0; each clip its own (see below)

A folder here holds at most 10,000 files, so fsd50k/clips/eval/ is split into ten folders named for the last digit of the clip's id (clips/eval/7/123457.wav). MMSB reads either layout.

Not here, and why

Six scenarios need a source that is not here. A seventh, v7_screen, needs none: the benchmark generates it.

Scenario Source Why it is not copied here Where to get it
a1_home_audio SINS Its licence allows academic research only, with no transfer to third parties mmsb fetch --scenario a1_home_audio, or Zenodo 2547315
v10_home_clips Charades Its licence does not grant the right to distribute the data to a third party mmsb fetch --scenario v10_home_clips, or its page
i2_screenshots Android in the Wild We found no stated licence for the data google-research/android_in_the_wild
v2_dashcam Australian Roads Dashcam Driving CC BY-NC 4.0, but the picture shows GPS readouts and number plates that cannot be removed without re-encoding it mmsb fetch --scenario v2_dashcam, or qutegocentric/Australian_Roads_Dashcam_Driving
i1_camera_roll ATM-Bench CC BY-NC 4.0, but it is one person's phone photos and each one carries GPS in its EXIF mmsb fetch --scenario i1_camera_roll, or Jingbiao/ATM-Bench
v8_egolife EgoLife MMSB leaves this download to you, and no copy was made for this repository lmms-lab/EgoLife; MMSB's MEDIA.md gives the folders

One file is also left out on purpose: BirdVox-full-night_gps-coordinates.csv, which says where the recorders stood. MMSB does not read it, and BirdVox's Zenodo record still has it.

Use

hf download DanMcInerney/mmmb-sources --repo-type dataset --local-dir D:/mmsb-sources
set MMSB_SOURCES=D:/mmsb-sources
uv run mmsb build --scenario i3_trailcam --scale standard --configs files,listing

To download only what one scenario reads, add --include with its folder from the table, such as --include "caltech-camera-traps/*".

The build copies the files it needs into a pile of its own, under the names a real device would give them, and sets each file's time to when it was recorded, so the download's own file times do not matter. Agents are refused any command that names the sources folder, so give it a name of 6 or more letters that no tool prints, such as mmsb-sources.

Where each part comes from

  • MEVA: the MEVA dataset (video from its public bucket, labels from meva-data-repo, whose LICENSE and README.md are kept here). Corona, K., Osterdahl, K., Collins, R. and Hoogs, A. "MEVA: A Large-Scale Multiview, Multimodal Video Dataset for Activity Detection." WACV 2021.
  • SoccerTrack v2 follow-cam clips: algo-dev99/soccertrack-v2-followcam-events, clips and labels by algo-dev99, from Scott et al., SoccerTrack v2.
  • AMI Meeting Corpus: groups.inf.ed.ac.uk/ami/corpus. Carletta et al. "The AMI Meeting Corpus: A Pre-announcement." MLMI 2005.
  • MIT OpenCourseWare: ocw.mit.edu. 14.01 Principles of Microeconomics, Fall 2023 (archive.org); 6.0002 Introduction to Computational Thinking and Data Science, Fall 2016 (archive.org); the podcast Chalk Radio (archive.org); and the course PDFs from each course's own page (5.111SC, 6.006, 7.016 and 9.00SC).
  • Caltech Camera Traps: LILA. Beery, S., Van Horn, G. and Perona, P. "Recognition in Terra Incognita." ECCV 2018.
  • BD3: chandrabhuma/building_defect_vqa.
  • SROIE: jsdnrs/ICDAR2019-SROIE. Huang et al., ICDAR 2019 Robust Reading Challenge on Scanned Receipts OCR and Information Extraction.
  • CORD v2: naver-clova-ix/cord-v2. Park et al. "CORD: A Consolidated Receipt Dataset for Post-OCR Parsing." Document Intelligence Workshop at NeurIPS 2019.
  • iNaturalist: inaturalist.org and its open-data bucket. Every photo is by iNaturalist user 5669063, who holds its copyright. observations.jsonl gives each photo's observation id and licence; https://www.inaturalist.org/observations/<id> is that photo's page and credit. The photos are for non-commercial use only.
  • ICSI Meeting Corpus: groups.inf.ed.ac.uk/ami/icsi. Janin et al. "The ICSI Meeting Corpus." ICASSP 2003.
  • BirdVox-full-night: Zenodo 1205569. Lostanlen, Salamon, Farnsworth, Kelling and Bello.
  • Free Music Archive: mdeff/fma. Defferrard, Benzi, Vandergheynst and Bresson. "FMA: A Dataset for Music Analysis." ISMIR 2017. The metadata is CC BY 4.0. The dataset's authors do not hold the copyright on the audio and distribute it under the licence each artist chose. tracks.csv in fma_metadata.zip names each track's artist, title and licence, which is the attribution. Most of fma_small is non-commercial, and many tracks allow no derivative works. For 5 of the 8,000 tracks that field is empty. The archives are kept whole so that a build matches one from the original download.
  • Robocall audio: wspr-ncsu/robocall-audio-dataset, NC State University; the calls are from FTC cases. Its README says the data is in the public domain and its LICENSE.md says CC BY-ND 4.0. The stricter reading is taken: the files here are unchanged, and a pile built from them is not for passing on.
  • FSD50K: Zenodo 4060432. Fonseca, E., Favory, X., Pons, J., Font, F. and Serra, X. "FSD50K: An Open Dataset of Human-Labeled Sound Events." IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2022. FSD50K as a whole is CC BY 4.0 (fsd50k/LICENSE-DATASET.md). Each clip keeps its Freesound licence: CC0, CC BY, CC BY-NC or CC Sampling+. fsd50k/metadata/eval_clips_info_FSD50K.json names each clip's licence and uploader, which is the attribution. Clips under CC BY-NC are for non-commercial use only.

What was changed

Every media file is as its source published it, with one exception: the photos and labels in bd3-building-defects/test_images/ and test_images_labels.jsonl were written out of the source's parquet file, which is also here. inat-observer/observations.jsonl was written from iNaturalist's records, and each MANIFEST.jsonl is mmsb fetch's own list of what it took, with each file's SHA-256.

For meva, caltech-camera-traps and fsd50k, a pile built from this copy was checked against one built from the original downloads: the same questions, and every file the same in name, size and time.

Downloads last month
91