Named Entity Recognition Model

Runs inside your own AWS account: your text is processed on a SageMaker endpoint you control, inside your VPC, and is never sent to a third-party API.

This SageMaker model package provides a REST API to detect named entities in US English text. Send a JSON array of sentences and receive labeled spans with character offsets and confidence scores.

Eighteen entity types are detected by default (person, organization, location, date, money, and more). Pass an optional "labels" array to extract any entity types you need (drugs, statutes, tickers) without retraining.

GPU is optional. Deploy on a GPU instance (for example ml.g4dn.xlarge) for higher throughput; the same image uses CUDA automatically when a GPU is present, and falls back to CPU otherwise. The API accepts JSON or CSV and supports real-time endpoints and SageMaker batch transform.

We welcome your feedback at [email protected]. Usage notebook: https://colab.research.google.com/drive/1iL1Q0FiYfKoUqYzglpeBFDxzDjV8IF9s

Use cases: redacting or routing documents by the people, companies and places they mention, enriching support tickets and contracts with structured fields, and feeding entity data to search, analytics and compliance workflows.

Known limitations: English text only. Event, facility and product are the least accurate default types, and very long inputs should be split into sentences or short paragraphs.

Model and training data

GLiNER medium (urchade/gliner_medium-v2.1), fine-tuned on a sample of Few-NERD (Ding et al., ACL-IJCNLP 2021) plus Sigmodata's own synthetic and hard-negative examples. The container runs in network isolation, so no data leaves your account.

Measured performance - Span-level micro F1 of 0.956 (precision 0.960, recall 0.952) on our validation set. An entity counts as correct only on an exact character-span and label match. - F1 by type: 0.97 or higher for person, location (GPE), date, time, money, percent, quantity, ordinal, cardinal and language; 0.90 to 0.96 for organization, group (NORP), law, work of art and non-GPE location; 0.84 to 0.86 for event, facility and product. - Dates, times, money, percentages, quantities, ordinals and numbers are extracted by deterministic rules rather than the model. - Throughput about 7 texts per second on ml.m5.xlarge at a batch size of 25.