Chrome extension¶

OpenDecider Focus filters your YouTube feed with opendecider-nano running on your computer. It hides the kinds of video you choose, or the topics you name in your own words, and shows the rest. The titles it reads are never sent anywhere: there is no server, no account and no API key, and after one download it works offline.

It also checks text for prompt injection: select text on any page, right-click it, and choose Check with OpenDecider guard.

Install¶

Until it is on the Chrome Web Store, install it from a release:

1. Download opendecider-focus-<version>.zip from the latest release and unzip it.
2. Open chrome://extensions, turn on Developer mode, click Load unpacked and pick the unzipped folder.
3. Click the OpenDecider Focus button in the toolbar and press Download the model (once: 755 MiB on a computer with a usable GPU, 569 MiB without one).

Each release's zip carries a build provenance attestation and a Sigstore signature: gh attestation verify opendecider-focus-<version>.zip --repo manjunathshiva/opendecider. It needs Chrome 124 or later, on a computer (Chrome on phones has no extensions).

What it filters¶

The popup has three filters; a video is hidden when any of them says so.

• Kinds of video. Each video is sorted into one of 11 kinds: news, how-tos and recipes, science and tech, lectures and explainers, business, interviews and podcasts, music, gaming, comedy, celebrity and TV, vlogs. Tick the kinds to hide. Focus (the default) hides music, gaming, comedy, celebrity and TV, and vlogs.
• Your own rule. "Hide videos about crypto, celebrity gossip", or "Show only videos about cooking". The topics become one yes/no question about each video ("Is this video about crypto, celebrity gossip?"); write a whole question ending in ? to ask it as written.
• Shorts. Hidden by their links, without the model.

How sure before showing a video moves the line between showing and hiding: towards "Show more", a video the model is unsure about stays; towards "Hide more", it goes. It filters the home page, search results and the list beside a video. New videos stay hidden until they are judged (a fraction of a second on a GPU, once the model is loaded), so nothing flashes past. Each video is judged once and the answers are kept on your computer: hiding other kinds applies at once, and a new rule asks each video once more.

How accurate is it¶

Measured on 400 YouTube videos (the test half of the youtube suite of the benchmarks), with the creator's own category as the label. The numbers are balanced accuracy: the mean of the share of wanted videos kept and of unwanted ones hidden.

The rules, written as a viewer might: "Is this a music video or a song?", "Is this video about video games or gameplay?", "Is this a news or current affairs video?".

• Your own rule is where it is strongest: within a point of Jev, TypeSafe's decision model in the cloud (ahead of it on music), and ahead of Laya, the other open decision model, on every rule.
• The kinds are harder, and there Jev leads by 8.5 points. Whether a cooking show is a how-to or entertainment is a judgement call, and the creator's category often disagrees with it (about one video in six, by our reading of Quietly's own keep and hide lists); opendecider-nano keeps more borderline videos than Jev does, so expect a few comedy and talk-show clips in the Focus feed.
• Quietly's own request does not work with it. Quietly asks Jev one question with the title inside the question and the same state every time; opendecider-nano judges the state, so it answers that request the same way for every video. OpenDecider Focus puts the video in the state instead.
• English titles. opendecider-nano was evaluated in English only, and so was this: titles in other languages and scripts were not measured.

Speed and memory¶

Times on an Apple M4 Max in Chrome (8 threads without the GPU); a slower computer takes longer.

The model is kept in the browser after the first download, and later loads in one or two seconds. It is loaded when you open YouTube and closed after 10 minutes without filtering, which gives the memory back. Without a GPU, the feed fills in a few videos at a time.

Privacy and permissions¶

What leaves your computer: nothing you watch or search. The extension downloads the model's files from Hugging Face once (opendecider-nano-ONNX, at the revision @opendecider/web pins), and checks each file's SHA-256 before using it. It makes no other network request (its pages' Content Security Policy allows only those hosts), and it has no analytics.

The guard check reads only the text you select and right-click, and shows its answer in a small window. The full privacy policy lists what the extension reads, keeps and downloads.

Limits¶

• It reads YouTube's pages. When YouTube changes how a feed is built, the filter can stop finding videos until the extension is updated: the feed is then shown unfiltered, and the popup counts no videos on the page.
• Home, search and the list beside a video. Channel pages, subscriptions and playlists are not filtered.
• English, as above, and a computer: Chrome on phones has no extensions.

OpenDecider Focus is not affiliated with or endorsed by YouTube or Google.

Build it yourself¶

(cd typescript && npm ci && npm run build)
(cd web && npm ci && npm run build)
cd extension && npm ci && npm run build     # then Load unpacked: extension/dist
npm test                                    # the logic
npx playwright install chromium && npm run e2e   # end to end, in a fresh Chromium profile

npm run zip builds the Web Store package; it rebuilds byte for byte from the same sources. After you rebuild, press the reload button on chrome://extensions: Chrome keeps an unpacked extension's old service worker otherwise.

The idea of a feed filter that judges each title with a decision model comes from Quietly (MIT), which asks TypeSafe's Jev in the cloud. OpenDecider Focus is written separately, and runs the model on your computer.