streamtv/video-pipeline

StreamTV TikTok Live Pipeline Release

Overview

StreamTV TikTok Live Pipeline Release is a dataset export containing TikTok Live recording sessions with remuxed MP4 video chunks, ClickHouse metadata, OpenAI ASR transcripts, event logs, and rule-based classifier labels.

• Size: 364 streams, 0 MP4 chunks, 2.1 GB of local video files
• Modality: Video + Audio + Text + Live event metadata
• Source: TikTok Live recordings captured by the StreamTV recording pipeline
• Release: release-5m-2h-2026-04-21
• Status: Release export

Schema

Core Files

Path	Type	Description
data/manifest/streams/part-*.parquet	Parquet	Selected shared streams included in this export
data/video_segments/part-*.parquet	Parquet	MP4 segment index with paths, durations, source segment counts, and source keys
videos/<room_id>/<rendition>-<part>.mp4	MP4	Remuxed video chunks generated from HLS source segments
data/transcriptions/openai_gpt/part-*.parquet	Parquet	OpenAI ASR transcript rows aligned to exported MP4 chunks
data/classifiers/stream_labels/part-*.parquet	Parquet	Rule-based stream label outputs with evidence
data/events/**/part-*.parquet	Parquet	Raw and normalized live event tables exported from ClickHouse
data/recording/**/part-*.parquet	Parquet	Recording session, source segment, gap, failure, and playlist metadata

Stream Manifest Columns

Column	Type	Description
room_id	string	TikTok Live room identifier
stream_id	string	Stream identifier when available from source metadata
unique_id	string	Creator unique identifier when available
recording_session_id	string	Recording pipeline session identifier
data_viewer_shared	bool	Share flag from recording.stream_access
shared_at	timestamp	When the stream was shared for export access
source_segment_count	int	Number of source HLS segments available for the session
source_duration_ms	int	Total available source media duration in milliseconds
release_name	string	Export release name

Video Segment Columns

Column	Type	Description
video_id	string	Exported MP4 identifier
room_id	string	TikTok Live room identifier
recording_session_id	string	Recording pipeline session identifier
rendition_id	string	Source rendition, usually source
part_index	int	Chunk index within the stream and rendition
output_path	string	Relative path to the MP4 file
segment_count	int	Number of source HLS .ts files used
duration_ms	int	Indexed MP4 duration in milliseconds
byte_count	int	MP4 byte size
source_s3_keys	JSON array	Source object keys used for remuxing

Transcription Columns

Column	Type	Description
room_id	string	TikTok Live room identifier
recording_session_id	string	Recording pipeline session identifier
video_id	string	Exported MP4 identifier
part_index	int	Chunk index within the stream and rendition
start_ms	int	Transcript start offset relative to the MP4
end_ms	int	Transcript end offset relative to the MP4
text	string	ASR-generated transcript text
language	string	Language code returned by the transcription provider when available
provider	string	Transcription provider
model	string	Transcription model name

Classifier Label Columns

Column	Type	Description
room_id	string	TikTok Live room identifier
recording_session_id	string	Recording pipeline session identifier
label	string	Stream label such as gaming, battle, ecommerce, or irl_candidate
confidence	float	Rule confidence score between 0 and 1
source	string	Evidence source used by the classifier
evidence	string	Compact evidence string for the label
provider	string	Label provider, currently rule-based
model	string	Model name or rule version
created_at	timestamp	Label creation time

Dataset Counts

Stage	Status
Metadata	complete
Video	pending
Transcriptions	pending
Classifiers	pending

Output	Count
Streams	364
MP4 files	0
Transcription shards	0
Transcription rows	0
Classifier labels	0
Classifier rows	0
Video segment rows	0
Source HLS segment rows	139231
Event rows	1321987

Collection Methodology

This export was produced by the StreamTV TikTok Live recording and exporter pipeline:

1. Selection: Shared streams were selected from recording.stream_access where data_viewer_shared = true, filtered by configured duration constraints and ordered by the latest share or session timestamp.
2. Metadata export: ClickHouse recording and event tables were exported to Parquet with raw warehouse identifiers preserved.
3. Video remux: Source HLS .ts objects were downloaded from object storage, converted into VOD manifests, and remuxed with ffmpeg using stream copy.
4. Transcription: Exported MP4 chunks were transcribed with OpenAI ASR and written as Parquet rows.
5. Classification: Initial stream labels were generated with rule-based classifiers using room metadata, live intro text, hashtags, shopping events, battle/link events, game tags, and transcript text when present.

Usage

Using Pandas

import pandas as pd

segments = pd.read_parquet(
    "hf://datasets/streamtv/video-pipeline/data/video_segments/part-000000.parquet",
)

print(segments[["room_id", "output_path", "duration_ms"]])

Accessing Videos

Videos are stored in the videos/ directory and linked from data/video_segments:

import pandas as pd

segments = pd.read_parquet(
    "hf://datasets/streamtv/video-pipeline/data/video_segments/part-000000.parquet",
)

video_path = segments.iloc[0]["output_path"]

Reading Transcripts and Labels

import pandas as pd

transcripts = pd.read_parquet(
    "hf://datasets/streamtv/video-pipeline/data/transcriptions/openai_gpt/part-000000.parquet",
)
labels = pd.read_parquet(
    "hf://datasets/streamtv/video-pipeline/data/classifiers/stream_labels/part-000000.parquet",
)

Notes

• Release staging: Metadata, video, transcripts, classifiers, manifests, and this README are published as staged release artifacts.
• Video chunks: The release materializes planned MP4 chunks from the selected streams, including shorter tail chunks and discontinuity islands when source media is available.
• Transcripts: ASR text is machine-generated and may contain errors, omissions, or language detection mistakes.
• Classifiers: Labels are rule-based unless the provider column indicates another provider.
• Identifiers: This first smoke export preserves raw room, stream, session, and source identifiers.
• Access controls: Raw event payloads can contain user-generated text and platform metadata. Use the dataset only under the access terms associated with the repository.

Larger Versions

Larger staged releases can use the same schema shape with additional streams, MP4 chunks, transcript shards, and classifier labels.

Citation

@dataset{streamtv_video_pipeline_2026,
  title={StreamTV TikTok Live Pipeline Release},
  author={StreamTV},
  year={2026},
  url={https://huggingface.co/datasets/streamtv/video-pipeline}
}

License & Disclaimer

This dataset is provided for research and experimental use under the repository access terms.

StreamTV does not claim ownership of the underlying video content. Users are responsible for ensuring that their use complies with applicable laws, platform terms, and repository restrictions.