Datasets:

katana-video
/

webcodecs-codec-support

license: cc-by-4.0
task_categories:
  - tabular-classification
tags:
  - webcodecs
  - video-processing
  - hardware-acceleration
  - telemetry
  - browser-support
  - av1
  - h265
  - hevc
  - vp9
  - h264
pretty_name: WebCodecs Codec Support Dataset
size_categories:
  - 1B<n<10B
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet

The Codec Support Dataset

Version 2026-10-01

Dataset Summary

This dataset contains real-world audio and video codec support data passively collected from 7,655,853 unique user sessions at free.upscaler.video, an open-source browser-based video upscaler with ~700,000 monthly active users. Each session uses the WebCodecs API (VideoEncoder.isConfigSupported() / VideoDecoder.isConfigSupported() and their audio equivalents) to test encoding and decoding support for a random sample of ~300 out of 1,087 codec strings across the major codec families (H.264, H.265, VP8, VP9, AV1, AAC, Opus, etc.).

Total Tests: 2,434,549,264
Sessions: 7,655,853
Unique Codecs: 1,087
Collection Period: January 2, 2026 – October 1, 2026
Format: Parquet (zstd), one file per month, 3.85 GB total
DOI (all versions): 10.5281/zenodo.19187466

The data is summarized interactively at webcodecsfundamentals.org, with a per-codec support table and an analysis of AV1, HEVC, VP9 and H.264 support.

Files

File	Days	Rows
`data/codec-support-2026-01.parquet`	30	138,596,578
`data/codec-support-2026-02.parquet`	27	81,443,298
`data/codec-support-2026-03.parquet`	31	201,598,644
`data/codec-support-2026-04.parquet`	30	197,135,832
`data/codec-support-2026-05.parquet`	31	236,605,038
`data/codec-support-2026-06.parquet`	30	281,177,508
`data/codec-support-2026-07.parquet`	31	343,646,700
`data/codec-support-2026-08.parquet`	31	454,035,630
`data/codec-support-2026-09.parquet`	30	484,965,900
`data/codec-support-2026-10.parquet`	1	15,344,136

SHA256SUMS contains checksums for every file.

Dataset Structure

One row per codec test (one codec string tested in one session). Rows from the same session are contiguous.

Column	Type	Description
`session_id`	string	Unique ID of the test session
`timestamp`	timestamp (ms, UTC)	When the session's tests were run
`user_agent`	string	Full browser user agent string
`browser`	string	Browser family (Chrome, Safari, Edge, Firefox, Unknown)
`platform_raw`	string	Raw `navigator.platform` value
`platform`	string	Normalized platform (Windows, macOS, iOS, Android, Linux, Other)
`codec`	string	WebCodecs codec string tested (e.g. `av01.0.16M.08`)
`encoder_supported`	bool	Whether the encoder supports this codec
`decoder_supported`	bool, nullable	Whether the decoder supports this codec; null if not tested

decoder_supported is null for sessions before January 14, 2026, when decoder testing was added (98% of sessions include decoder data). browser is reported by the client; all Chromium-based browsers other than Edge (Brave, Opera, Samsung Internet, etc.) are grouped as Chrome.

Platform Normalization

Win32, Win64, WIN32 ➔ Windows
MacIntel, MacOS, Mac OS ➔ macOS
iPhone, iPad ➔ iOS
Linux armv7l, Linux armv8l, Linux armv81, Linux armv6l, Linux arm, Android, Android64 ➔ Android
Linux, Linux x86_64, Linux aarch64, Linux amd64, Linux i686 ➔ Linux
Anything else (e.g. HarmonyOS, blank, spoofed values) ➔ Other (18 sessions)

Usage

At 2.4 billion rows the dataset is too large to load into pandas; query the files in place with DuckDB, Polars or Spark:

import duckdb

con = duckdb.connect()
con.sql("CREATE VIEW tests AS SELECT * FROM read_parquet('data/*.parquet')")

# Encoder and decoder support by codec
con.sql("""
    SELECT codec,
           round(100 * avg(encoder_supported::INT), 2) AS encoder_pct,
           round(100 * avg(decoder_supported::INT), 2) AS decoder_pct,
           count(*) AS tests
    FROM tests
    GROUP BY codec
    ORDER BY encoder_pct DESC
""").show()

With Hugging Face datasets, use streaming rather than downloading everything up front:

from datasets import load_dataset

ds = load_dataset("katana-video/webcodecs-codec-support", split="train", streaming=True)

Dataset Creation & Methodology

Data is collected anonymously in the background on free.upscaler.video. No personally identifiable information is collected: only the user agent, navigator.platform, and codec test results. A full description of the collection methodology is here.

Known Limitations

Sessions, not devices. Percentages are per session and reflect free.upscaler.video's traffic mix (77% Chrome, 50% Windows, 30% Android). Reweight by your own traffic for other applications.
WebCodecs API support, not hardware support. A codec the device supports may be missing from the WebCodecs API (e.g. Firefox on Android does not expose WebCodecs), so the numbers are conservative.
isConfigSupported() only. Results do not measure encode/decode performance or quality.
Collection gap. Very few sessions were recorded between February 8 and February 20, 2026 (none on February 19) due to a collection outage.

Changes from Version 2026-03-22

7,655,853 sessions (was 1,142,586), collection period extended to October 1, 2026
Parquet instead of CSV (3.85 GB instead of 69.4 GB uncompressed)
New session_id column
Platform normalization extended to rare navigator.platform values, with an Other bucket
Producer: Katana Video Inc.
Author: Samrat Bhattacharyya

Citation Information

@dataset{upscaler_codec_dataset_2026,
  title        = {The Codec Support Dataset},
  author       = {Bhattacharyya, Samrat},
  year         = {2026},
  version      = {2026-10-01},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.19187466},
  url          = {https://doi.org/10.5281/zenodo.19187466},
  note         = {2.43B codec tests from 7.66M sessions}
}

License

Creative Commons Attribution 4.0 International (CC-BY 4.0). Credit "The Codec Support Dataset" with a link to webcodecsfundamentals.org/datasets/codec-support.