Wikipedia:Edit filter noticeboard

Emoji vandalism missed by emoji filter

Special:Diff/1325372991 at Issam Nabahin (now RD3'd) used colored block emojis, which are not in the Special:AbuseFilter/680, to make a Pepe the Frog ASCII art. These are U+2B1B–2B1C (⬛⬜) and U+1F7E5–1F7EB (🟥🟦🟧🟨🟩🟪🟫). A use of the emoji version of U+26A0 ⚠️ WARNING SIGN was also missed. –LaundryPizza03 (dc̄) 21:27, 3 December 2025 (UTC)

Also (in order of Unicode point): ▪️▫️◻️◼️◽️◾️⚪️⚫️❤️💙💚💛💜🔴🔵🔶🔷🔸🔹🔺🔻🟠🟡🟢🟣🟤🤍🤎🧡🩵🩶🩷, and possibly 🔲🔳. –LaundryPizza03 (dc̄) 21:38, 3 December 2025 (UTC)

No, wait, probably the filter did catch the ⚠️, but not the colored blocks. –LaundryPizza03 (dc̄) 06:59, 5 December 2025 (UTC)

I've edited the filter to add the two ranges U+2B1B–2B1C and U+1F7E5–1F7EB. We should probably think about widening these ranges; for example, the latter range belongs to the block Geometric Shapes Extended, which consists entirely of wingdings-like characters. — The Anome (talk) 13:55, 15 December 2025 (UTC)

Perhaps we could start with readjusting the current blocks for Emoji; if done correctly, there should be no outliers (e.g. U+1F6FA 🛺 AUTO RICKSHAW is not in the current filter). There could also be a rule for U+FE0E and U+FE0F (e.g. ⏳ → ⏳︎ and ⚠ → ⚠️), which are both useless for anything but emoji and avoids unnecessary hassle from trying to deal with emoji variants of actually useful characters such as © (→ ©️). –LaundryPizza03 (dc̄) 17:26, 15 December 2025 (UTC)

I'm working on this at User:LaundryPizza03/filter 680. –LaundryPizza03 (dc̄) 18:57, 15 December 2025 (UTC)

Hey there. I just reverted , which is the second time I've reverted an emoji this week (I don't know when the first time was). Just thought I would let you guys know. win8x (talk) 23:08, 17 December 2025 (UTC)

Testing on 110

I'm going to disable this once it hits 5,000 matches. I'll follow-up on the mailing list or here if I end up making this filter public (which seems likely at this point). Daniel Quinlan (talk) 23:26, 12 December 2025 (UTC)

I ended up disabling it earlier due to the insane false positive rate. I initially marked it as private since there was an outside chance it might be useful in an LTA-focused filter (depending on the matches), but it's so far from being useful that keeping it private served no purpose. See below for the results of the test. Daniel Quinlan (talk) 00:17, 13 December 2025 (UTC)

Do not use revertrisk_level

I did a brief test of revertrisk_level, a relatively new variable added to the abuse filter, at 110 (hist · log). There's some documentation on the feature at:

• mw:Extension:ORES/AbuseFilter variables
• mw:2025 RecentChanges Language Agnostic Revert Risk Filtering
• meta:Machine learning models/Production/Language-agnostic revert risk

I was concerned about the high false positive rate reported in the second link, but I was hoping that it might be usable as a throttle-based filter. Unfortunately, it looks like this is completely unusable and should not be used in any filters at this time:

• The variable is very expensive with a reported run time of about 170 ms to 175 ms.
• The hit rate was extraordinarily high, matching about 43% of all actions (it only matches on edits so the hit rate on edits is even higher). The vast majority of the hits look like false positives.

The test was simply looking for a revertrisk_level of "high". Unfortunately, it is primarily flagging routine edits like adding templates and categories. This suggests that the model isn't trained properly although it could be some other problem. I've reached out to the WMF. Daniel Quinlan (talk) 00:08, 13 December 2025 (UTC)

Thanks for the test. If 170 ms is the average, I wonder what the worst case is. Suffusion of Yellow (talk) 22:32, 14 December 2025 (UTC)

Edit distance/fuzzy matching

We have an LTA that is trying to get round filters by just changing a few letters of their chosen identity. Specifically, getting around Special:AbuseFilter/1395. Do we have some sort of edit distance-based metric or fuzzy matching algorithm we can use here? Ccnorm is not sufficient. — The Anome (talk) 22:47, 14 December 2025 (UTC)

Unfortunately, not yet; see phab:T274062. Suffusion of Yellow (talk) 23:42, 14 December 2025 (UTC)

That's annoying. The trigram version of that would be very effective in this case. I guess I can code up a program to produce a huge brute-force set of regexes that might approximate this somehow, but that's not a good use of anyone's resources. — The Anome (talk) 14:05, 15 December 2025 (UTC)

Set 1389 (hist · log) to tag or warn

1389, the filter that logs edits where someone removes a declined block request, has been working well for over a month now. I would like to get everyone's opinions before setting to tag or warn (if we warn, we would have to also create a custom warning message of course). Courtesy ping to @Pppery: for initially requesting this filter at EFR. – PharyngealImplosive7 (talk) 14:49, 18 December 2025 (UTC)

How is the false negative rate? I would actually be in favor of settings this to tag-only then doing away with the WP:KEEPDECLINEDUNBLOCK rule entirely, or at least softening it to something like "you must restore any declined unblock requests before requesting again." It would then be the responsibility of the reviewing admin to check for any tagged edits. The unblock template could even include a convenience link to the tagged edits, wrapped in sysop-show.

WP:KEEPDECLINEDUNBLOCK is frankly a punitive rule and even if we keep it, we should not be creating more technical measures to enforce it. Accept your fate? Great, blank your talk page per WP:NOTWALLOFSHAME and move on. But dare to question the block? Now there's a page, possibly titled after your real name (or a name easily traceable to you), that's up here forever. Where someone calls you "incompetent". For something you did ten years ago, when you were twelve. At least a human might recognize this, and not revert per IAR. But a filter cannot. Suffusion of Yellow (talk) 19:18, 21 December 2025 (UTC)

I'm not sure about the exact false negative rate (because I haven't checked recent changes to monitor for any missed edits). Regardless, the false negative rate still shouldn't be too high because of how the filter logic is structured (unless I'm missing something). As for the policy, I understand the concern, which I guess might be why tagging is preferable to a stronger action? – PharyngealImplosive7 (talk) 01:05, 22 December 2025 (UTC)