Ertas for Named Entity Recognition

The Challenge

Named entity recognition is the foundation of structured information extraction from unstructured text. Generic NER models can identify common entities like person names, locations, and organizations, but they fall apart on domain-specific entity types. A medical NER system needs to recognize drug names, dosage units, anatomical terms, and ICD codes. A legal system needs to identify case citations, statute references, party names, and jurisdictions. A financial system needs to extract ticker symbols, monetary amounts with currency, regulatory body names, and specific financial instruments.

Building custom NER models has traditionally required deep NLP expertise, complex annotation tooling, and significant engineering effort to deploy and maintain. Teams spend months on annotation guidelines, inter-annotator agreement measurement, and model architecture selection before they even begin training. The result is often a fragile pipeline that breaks when encountering entity formats it was not explicitly trained on, such as a new drug naming convention, an unusual citation format, or a foreign organization name that does not match expected patterns.

The Solution

Ertas simplifies custom NER by building on the broad language understanding of open language models and focusing fine-tuning on the entity patterns specific to your domain. Instead of training a traditional NER model from scratch, teams fine-tune a generative model in Ertas Studio on examples of text paired with the entities it contains, written in a structured output format such as JSON. The model learns to find entities from the context around them as well as from their surface form.

Because the base model already understands language well, the fine-tuned model tends to handle entity formats it has not seen explicitly, such as a new drug name in a familiar sentence position or an unusual organization name in a role organizations usually fill. Teams should confirm this on their own held-out examples before relying on it. Studio accepts training data as JSONL with input text and structured entity output, and Data Craft, Studio's guided dataset builder, lets annotators edit and approve examples in one place. The trained model is exported from Hub as GGUF or safetensors and runs on your own servers, via Ollama for batch jobs or vLLM for real-time extraction, so the text being processed stays inside your network. Training runs on Ertas's cloud GPUs and the annotated examples are uploaded to Ertas, so where they contain personal data, replace real names and identifiers with realistic stand-ins before upload.

Key Features

Example Workflow

A pharmaceutical company needs to extract drug names, dosage information, adverse events and patient demographics from clinical trial reports for pharmacovigilance monitoring. The NLP team annotates 15,000 report excerpts with its custom entity schema (12 entity types), replaces patient identifiers with synthetic stand-ins, and uploads the JSONL dataset to Ertas Studio. Annotators settle disputed examples in Data Craft before training. The team fine-tunes Qwen 2.5 7B to take report text as input and return structured JSON with every entity, its type and its text span, then measures F1 per entity type on a held-out set of annotated reports, alongside the generic NER system it used before. The model is exported from Hub and runs as a nightly batch job on the company's own servers over newly received trial reports. Extracted entities load into the pharmacovigilance database, where safety analysts review flagged adverse events, and their corrections feed the next training run.