Local LLM Comment Moderation — A Model Swap Story

Comments Were Piling Up

When someone leaves a comment on my café’s guestbook, a human doesn’t review it first — a local LLM does. It decides whether the comment is abusive, an ad, or just a nice review, and handles it automatically.

https://cafepurplemint.com (a tiny bit of shameless promotion)

Then one day, while going through the logs, I noticed something odd.

[MOD] Processing 5 pending entries...
[MOD] Processing 5 pending entries...
[MOD] Processing 5 pending entries...

The same log line kept repeating. The next poll was kicking off before the previous batch had finished processing.

The cause was simple. Moderating a single comment took 5.6 seconds on average. The polling interval was 1 second. Mathematically, even a small burst of comments was guaranteed to back up the processing queue. (I’d actually just built the café’s homepage and — not really an ad, but kind of an ad — asked a group chat of friends to try leaving a comment.)

Something Was Off With the Prompt From the Start

Before swapping models, I first took a hard look at the prompt.

The original approach looked like this:

Judge pass: false if the comment matches any of the following criteria:
- Profanity or abusive language
- Advertising, spam, or external links
- Defamation of a specific person or business, or false claims
- Personal information (phone number, address, email, etc.)
- Hate speech or discriminatory remarks

It was a denylist-style approach. But given what a café guestbook actually is, the range of comments that should get through is much narrower than that. Visit reviews, compliments, thank-you notes. That’s basically it.

So I flipped the policy.

❌ Before: "Reject these" (denylist)
✅ After:  "Only allow these" (allowlist)

I also broke the categories down further.

meaningless is a category I added this time. Comments like lolololol, abcde, or just emoji. Not bad, not good — just meaningless. Auto-rejecting felt a bit unfair, so those get routed to the admin instead.

I also added a line at the end of the prompt explicitly listing the valid category values. While testing smaller models, I noticed some of them were returning out-of-spec values like "offensive".

category must be exactly one of the following:
"clean", "profanity", "spam", "advertising", "defamation", "negative", "meaningless", "other"

Lining Up Three Models

Once the prompt was cleaned up, I tested three candidate models against the same 8 cases. I hit each one directly with curl and compared response time, accuracy, and token counts.

Accuracy

qwen2.5:3b classified “What’s the restroom door code?” as profanity. As in, abusive language. That model is simply unusable.

Speed and Tokens

The smaller models are 4–5x faster. But they get it wrong. In particular, advertising slipped through both of them.

I Tried a Two-Stage Pipeline, But…

“If the small model is fast, why not use it as a first pass, and have gpt-oss re-check anything it flags?”

It sounded reasonable. First, I checked the false-positive rate on clean — if a small model wrongly rejects good comments, the whole approach is dead on arrival.

I fed it 10 normal comments, and both gemma3:4b and qwen2.5:3b came back with 0% false positives. A good sign.

But the opposite direction was the problem. If a small model waves a bad comment through as clean, it gets published without ever passing through gpt-oss. I measured this “miss risk.”

qwen2.5 did better, but advertising was a case neither model could reliably catch, no matter how I varied the prompt. I tried 7 different prompt variants — every single one failed. It seemed like 3–4B models simply couldn’t grasp the concept of “a post promoting some business other than this café.”

I dropped the two-stage pipeline idea.

gemma3:12b Solved Everything

Last, I tested gemma3:12b-it-qat. Being a 12B model, I expected it to beat the 4B version — but the result was far better than I’d anticipated.

It nailed all 12 test cases, including 5 advertising patterns, using the base prompt with no modifications at all.

✅ clean         → clean         (normal review)
✅ profanity     → profanity     (abuse)
✅ spam          → spam          (spam link)
✅ advertising   → advertising   (promoting another business) ← what 4b couldn't catch
✅ defamation    → defamation
✅ negative      → negative      (negative opinion)
✅ meaningless   → meaningless   (lolol)
✅ other         → other         (restroom code)
✅ advertising   → advertising   (includes an @social-account handle)
✅ advertising   → advertising   ("come visit my shop too")
✅ advertising   → advertising   (external link)
✅ advertising   → advertising   (implicit promotion)

And it’s fast.

It’s 45% faster than the old model, more accurate, and uses fewer tokens. No need for a complicated two-stage pipeline either.

Comparing the gemma3 Model Family

It was interesting to see how big the gap was between 4B and 12B within the same gemma3 family.

No matter how much I tuned the prompt, the 4B model never reliably caught advertising. The 12B model got it right out of the box, with the base prompt. This test made it very clear that the gap between model sizes isn’t just about speed — it’s a difference in the model’s level of conceptual understanding.

Conclusion — One Line in .env

No two-stage pipeline, no prompt tuning needed.

# before
LLM_MODEL=gpt-oss:20b

# after
LLM_MODEL=gemma3:12b-it-qat

One line changed.

Response time: 5.6s → 3.1s. Accuracy: 8/8 → 12/12. And the comment queue backing up? Gone.

This confirmed once again that with local LLMs, model choice is basically everything. There’s a limit to what prompt engineering can do — if the model doesn’t grasp the underlying concept, no amount of prompting will get you there. Once I hit that wall, the answer turned out to be simpler than expected: a bigger model.

There’s still room to improve on the trickier advertising cases — implicit promotion like “this place is nice, but you should also check out my shop…” — probably with a rule-based pre-filter. But that’s a story for another day. For now, this is good enough.

See also

• How AI Helped Me Clean Up a Cafe Menu
• Claude Code Writing Code Before You've Even Defined Requirements?