fastino/gliguard-LLMGuardrails-300M
Primitive: /extract · Extract ·
DeBERTa
GLiGuard is a compact, encoder-based guardrail model for LLM safety moderation built on the `GLiNER2` interface. Instead of generating moderation verdicts autoregressively, it treats safety as structured classification: you provide task names and candidate labels at inference time, and the model ...
View on Hugging Face → Fine-tuned from fastino/gliner2-base-v1
Overview
Hardware: — drives latency, throughput & cost
| Size | 208M params |
|---|---|
| Tasks | /extract |
| License | apache-2.0 |
| Languages | en |
| Latency | 55 ms |
| Throughput | 47.5K tok/s |
| Cost | $0.0047 /1M tok |
Cost is approximate — computed from list GPU prices; your actual price depends on the provider you deploy SIE with.
Extraction
| Output kinds | Class Labels |
|---|---|
| Inputs | text |
| Max sequence length | 512 |
Benchmarks
toxic-chat
Toxicity classification of real user-AI conversations (ToxicChat)
Corpus: 2,853 Queries: 2,853
Quality
f1 0.5756
precision 0.4154
recall 0.9365
accuracy 0.9016
Performance L4 b1 c16
Extract 47.5K tok/s
Extract p50 55.2ms
Compare (0)Compare →