Vela PII
Vela PII finds sensitive entity spans for privacy-aware routing and redaction.
307M parameters · Input capacity: 32,768 tokens, including special tokens.
Outputs use 35 BIO labels across 17 entity types. The example returns spans with Unicode character offsets.
Evaluation
Exact-span micro F1 (×100) on the same synthetic development sets, compared with the original mmBERT32K PII model. Higher is better.
| Evaluation | Original mmBERT | Vela |
|---|---|---|
| Short inputs · 888 | 23.66 | 90.76 |
| Controlled 4K context · 30 | 0.44 | 89.03 |
| Controlled 8K context · 30 | 0.28 | 89.88 |
| Controlled 16K context · 30 | 0.36 | 89.73 |
| Controlled 32K context · 30 | 0.23 | 89.24 |
Synthetic examples cover six languages. Long inputs include sparse entities, densely packed repeated entities and negative examples; micro F1 weights each entity equally. Both models process complete inputs in FP32 with the same exact-span scorer. These development sets informed Vela selection; they are not an independent natural-document benchmark.
Quick start
With PyTorch and Transformers 4.57.6 or 5.17.0:
from transformers import pipeline
model_id = "vllm-sr/Vela-1.0-Encoder-307M-PII"
model = pipeline("token-classification", model=model_id, aggregation_strategy="simple", device=-1)
text = "Contact Mara Wells at mara.wells@example.com."
assert len(model.tokenizer.encode(text)) <= model.model.config.max_position_embeddings
print(model(text))