knowledgator/gliclass-large-v3.0
Primitive: /extract · Extract ·
DeBERTa
This is an efficient zero-shot classifier inspired by GLiNER work. It demonstrates the same performance as a cross-encoder while being more compute-efficient because classification is done at a single forward path.
Overview
Hardware: — drives latency, throughput & cost
| Size | 439M params |
|---|---|
| Tasks | /extract |
| License | apache-2.0 |
| Latency | 94 ms |
| Throughput | 27.2K tok/s |
| Cost | $0.0082 /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
AG News
Topic classification of news articles into world, sports, business, and sci/tech categories
Corpus: 7,600 Queries: 7,600
Quality
accuracy 0.7429
Performance L4 b1 c16
Extract 27.2K tok/s
Extract p50 93.8ms
medical_questions_pairs
Classify whether two medical questions ask the same thing (question-pair similarity)
Corpus: 3,048 Queries: 3,048
Quality
accuracy 0.4820
Compare (0)Compare →