← sources

I used local Gemma 4 to moderate a community and it was it is promising

source: reddit (reddit:LocalLLaMA)published 2026-11-15 13:35ingested 2026-11-16 02:38

retained text

I run a couple of small communities and it's becoming difficult to moderate them all. So today I built a script on top of my existing app which I use to track the content on those communities.

The script does a few things based on where I am. So I store the general notes of my communities in a dedicated place I can easily query.

And in the moderation panel I go through each of the item which needs moderation and I run them through my notes and let Gemma 4 decide whether the item is following community notes or not. If Gemma thinks it's not, then my script will remove the content from public and add a reason as well.

At first I had tried GPT oss 20b model which was really faster but the results were not really good. Then I ran the same process through Gemma 4 with thinking enabled. Results were good but speed was extremely slow. So I tried with thinking disabled and I did not see any noticable drop in the results.

Now soon I will try to run the same thing at scale and see if the results are generally accurate and if they are, I might ship this tool with my friends.

I agree, the setup is quite tough and I've managed to run it properly on my network. If I have to share this with other people I may need to host the model somewhere and check if it can be scaled to a bigger number of requests per hour.

Anyways it was a really good experience. Especially when I had to build it on top of my existing app.

Just thought I'd share this small win over here

document fetched 2026-11-16 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1xwvnfr/i_used_local_gemma_4_to_moderate_a_community_and