⊘
Edge content moderation
SecurityReal-time moderation + PII redaction on user content — flag toxicity, self-harm, sexual content, violence and PII, fast enough to run inline on every message.
The routing recipe
Job typeClassification / routingPolicyFastest that passesConstraintscloudflare/groq
routes to
GPT-OSS 20B groq
groq:openai/gpt-oss-20b
89.3
match score
84.6
task fit
66.4
cost
96.4
speed
Why this model from benchmarks & capabilities
- "Classification / routing" leans hardest on instruction; GPT-OSS 20B scores 78/100 there — #37 of 54.
- Strength on this class of work shows up on IFEval, MMLU (zero-shot), where OpenAI (open-weight) models are competitive.
- Cost: ~$0.244/Mtok blended ($0.07 in / $0.30 out). Speed: ~1000 tok/s on Groq LPU.
- Chosen over Llama 3.1 8B Instant for higher task-fit (84.6 vs 75.6) & faster.
| Capability the job needs | Weight | GPT-OSS 20B | Catalog rank |
|---|---|---|---|
| instruction | 0.90 | #37 of 54 | |
| speed | 0.90 | #2 of 54 | |
| knowledge | 0.30 | #40 of 54 |
Relevant benchmarksIFEval MMLU (zero-shot)
EconomicsCost: ~$0.244/Mtok blended ($0.07 in / $0.30 out).
Speed: ~1000 tok/s on Groq LPU. vs runner-upChosen over Llama 3.1 8B Instant for higher task-fit (84.6 vs 75.6) & faster. (Llama 3.1 8B Instant, score 88.3)
Speed: ~1000 tok/s on Groq LPU. vs runner-upChosen over Llama 3.1 8B Instant for higher task-fit (84.6 vs 75.6) & faster. (Llama 3.1 8B Instant, score 88.3)
Example result
Sample input ▾
Hey call me at 415-555-0199 or email [email protected] — and tell that idiot in billing he is useless.
allow: false
categories: [harassment, pii]
severity: low
pii_found: [phone "415-555-0199", email "[email protected]"]
redacted: "Hey call me at [REDACTED] or email [REDACTED] — and tell that idiot in billing he is useless."
note: mild harassment ("idiot/useless") + contact PII; redact and allow with a soft warning, or route to human review per policy.
Test it on your own data