By kuben-developer · Presented by M90

TikTok videos / 4.5B

A research-scale dataset of captions, sounds, and engagement. Download the files. Start with a question.

ParquetZstandardResearch useMetadata only
4,501,811,789post records
289.47 GBcompressed files
27downloadable parts
16data columns

Files

27 parts · source revision cbcda8f

Each file is about 10.7 GB. Download one part or use the resumable downloader for the full release. Downloads use the M90 mirror when ready, with the original source as fallback.

Dataset card

Know what you’re working with.

One row per post. Captions, posting times, sounds, and engagement metrics collected over approximately three weeks. A sample of collected public posts, not a complete archive of TikTok.

Included

Post IDs, captions, engagement counts, sound identifiers, timing, and platform labels.

Not included

Video files, media URLs, author IDs, usernames, creator profiles, email fields, or comment text.

Engagement is a snapshot, not a time series. Adjacent rows may share a creator, but there is no creator join key. Record counts are publisher-reported.

View the 16-column schema +
content_iduint64

Unique post identifier.

create_timetimestamp

Publication time.

descstring

Original caption.

mentionslist<uint64>

Tagged accounts; not the author.

durationuint16

Duration in seconds.

is_videouint8

Video or photo-post flag.

music_iduint64

Sound identifier.

music_titlestring

Sound name.

views · likes · comments · shares · savesuint64

Engagement counts at collection time.

country · languagestring

Platform labels, not verified.

is_aduint8

Sponsored-post flag.

Start with less. Learn more.

Three ways to put data to work.

Book a 15-minute call
01

Query a focused sample.

Read only the columns you need. Explore captions, sounds, or engagement distributions before committing to a full scan.

SELECT music_id, COUNT(*)
FROM 'videos-00.parquet'
GROUP BY music_id;
02

Keep it in the cloud.

Use a cloud machine or a streaming object-storage pipeline. Preserve compression and verify checksums.

The download script writes to the machine where you run it. Choose a cloud volume with enough space.

03

Plan the video analysis.

Metadata can guide a research sample. Actual video retrieval is a separate step, subject to availability, API inputs, and usage rights.

Commercial workflows require appropriately licensed data. Book a call to scope the pipeline.

M90 in practice

What a data-led workflow can do.

See the full case studies ↗

Existing M90 case studies, as reported on our introduction page. These results are not attributed to this TikTok dataset.

Nina Clothing

159K views

Typical videos: 1K–3K views. A data-backed creative reached 159K.

See how it worked

Another account

2.5M views

The data-backed video reached 2.5M views. See the account comparison in the full case study.

Explore the workflow

Bring your question.

We’ll help build the pipeline.

From choosing the data to cloud storage and video analysis.
A focused 15-minute conversation about what you want to build.

Book a call with M90

Source & usage

Dataset collected and published by kuben-developer. M90 provides this directory and mirror with the publisher’s permission. The dataset retains its research and educational use terms. Do not use it to identify, profile, target, or contact individuals. Collection was contrary to TikTok’s terms. Neither M90 nor this dataset is affiliated with TikTok, ByteDance, or Hugging Face.

Original terms ↗Source correction or removal ↗Mirror removal request ↗