Backstage · dataset internals

deeper stats behind Telegram Code, measured from the lake catalog

internal view · numbers are the unbiased picture, filters and caveats included

·Viewers & pages Cmd/Ctrl+Shift+L from any page

1The size reality, half the "repos" are tiny

On an unbiased 2% sample (no filtering), the median source repo has just 7 code files and the distribution is dominated by a tiny-repo mass. This is why the downloadable dump filters to >6 files, but the honest distribution below keeps everything.

~48% of source repos have ≤6 code files. Of those, nearly half are single-file repos (one script, median 4.5 KB).

2What the single-file repos actually are

The ~49,600 single-file repos are not "projects", they are shared scripts. The most common filenames give it away: game cheats, Minecraft mods, Telegram bots, and generic entry points. This is a Telegram lake, and it shows.

3Language distribution (unbiased, all repos)

Scripting languages dominate the repo count because single-file scripts skew that way (JS, Python, Lua). By lines the ranking differs, C/C++ carries far more code per repo.

4Decompiled APK code, the bucket split

Per-file classification over a 2% APK sample (path-based; cross-app recurrence not applied here). By bytes:

The median app has 1.45 MB authored code to 2 KB obfuscated, obfuscation is concentrated in a heavy-tail minority, not the typical app. (The public site's line-share split additionally reclassifies cross-app-recurring files as shared library, which shifts "authored" lower.)

5Why the download is a subset, storage reality

Every statistic here is exact from the catalog, which covers all repos. But fetching complete repos (with config/docs) requires the raw zips, and those are mostly cold:

So the complete-repo download is drawn from the hot subset (recent uploads) and is not a representative sample, it is a quality-weighted showcase. The distribution numbers above never depend on it.

6Dedup tiers & containers