Datasets:

leee99
/

yt8m-h264

datasets:
  - name: leee99/yt8m-h264
language:
  - en
tags:
  - video
  - h264
  - bytestream
  - compressed-domain
  - youtube8m
  - machine-learning
  - dataset
license: mit
size_categories:
  - n<1K
task_categories:
  - other

📦 yt8m-h264

yt8m-h264 is a lightweight derivative of the YouTube-8M dataset that exposes H.264 NAL units (SPS, PPS, IDR) for efficient bytestream-level modeling and compression-aware video research.

This dataset stores pre-processed H.264 bytestream chunks directly in Arrow artifacts, allowing fast loading with:

from datasets import load_dataset
ds = load_dataset("leee99/yt8m-h264")

No decoding, FFMPEG, or custom scripts are required at load time.

✅ What’s inside?

Extracted from YouTube-8M video segments
Each sample contains:
- sps: Sequence of SPS NAL units (as raw bytes)
- pps: Sequence of PPS NAL units (as raw bytes)
- idr: Sequence of IDR slice NAL units (as raw bytes)
Stored directly in Arrow (binary) columns

A single example looks like:

{
    "sample_id": "00001234",
    "sps": [b"\x00\x00\x00...\x67"],          # list of byte payloads
    "pps": [b"\x00\x00\x00...\x68"],
    "idr": [b"\x65\x88\x99..."],              # IDR slices as raw byte arrays
}

These bytes correspond to Annex-B NAL units (0x00 00 00 01 <nal-header> <payload>), suitable for:

bytestream modeling
compressed-domain video understanding
tokenization (Byte-level / Bit-level)
entropy analysis
H.264 syntax learning

✅ Loading the dataset

from datasets import load_dataset
ds = load_dataset("leee99/yt8m-h264")

print(ds)
print(ds["test"][0])

Outputs:

DatasetDict({
    test: Dataset({
        features: ['sample_id', 'sps', 'pps', 'idr'],
        num_rows: <N>
    })
})