I'm seeing this on a bunch of random pages too. For example, https://omahaproxy.appspot.com/all offers to translate from Welsh. I see languages other than French on code review pages.

I'm bumping to a 1 as this is annoying.

Looks like "ios" is a Welsh word, but given the sheer number of non-Welsh words there...yea, seems like we're being extremely overaggressive.

Hm. Not happening with 54.0.2815.0 OSX

Would you mind looking at chrome://translate-internals/#detection-logs ?

(It updates after you visit the page in question, so open internals, visit codereview, go back to internals page and look at details)

+abakalov, who migrated us from CLD2 to CLD3 in 2827.0, and attaching my pre-migration detection logs. 

54.0.2838.0 contains a change on which text is used for detection as well - maybe that's causing issues here. (Also, omahaproxy should *really* serve a <meta name="google" value="notranslate"> :)

No longer happening to me for the original URL, but it's happening for me at https://codereview.chromium.org/2278823003/diff/20001/chrome/test/data/webui/net_internals/log_view_painter.js?context=25&column_width=80&tab_spaces=8 (Same review, another file).

Attaching the log file for that URL.

Deleted: translate_log.txt 31.6 KB

translate_log.txt 31.6 KB View Download

For whatever reason I lowered to 0, sorry about that.
Happens for me on https://codereview.chromium.org/2270543003/diff/240001/chrome/browser/ui/views/desktop_capture/desktop_media_source_view.cc I've attached the log file

Deleted: translate_internals_detect_logs_dump (1).json 10.8 KB

translate_internals_detect_logs_dump (1).json 10.8 KB View Download

Oops, I missed the dump button.  If you need another dump, just ask, and I'll attach one.

Thanks for reporting this issue! 
riesa@ and I are looking into it.

We submitted a CL on Friday that resolves this issue. The new predictions for the pages in this bug are now correct. The CL finds snippets of input text that are unreliable (e.g., repeated chunks, a lot of spaces), so that they are not considered during language prediction. I tried the change on the files in 10 or so randomly picked CLs as well.

We are trying to get this CL cherry-picked into 54.