UniversalNER offers a solution for open Named Entity Recognition (NER) by employing targeted distillation from large language models (LLMs). It aims to provide robust NER capabilities across diverse domains, benefiting researchers and developers working with natural language processing tasks.

How It Works

The project leverages targeted distillation, a technique that selectively transfers knowledge from powerful LLMs to a more specialized model. This approach aims to achieve high performance in NER while being more efficient than using a full LLM for every inference. The core architecture likely involves a transformer-based model fine-tuned on a curated dataset derived from LLM outputs.

Quick Start & Requirements

• Install via pip install -r requirements.txt after cloning the repository.
• Requires gcc version 5+ and CUDA versions between 11.0 and 11.8.
• Inference can be run on a single V100 16G GPU.
• Demo: https://github.com/universal-ner/universal-ner (Project Page)
• Inference code: https://github.com/universal-ner/universal-ner (Model)

Highlighted Details

• Offers both vLLM and Huggingface Transformers for inference.
• Provides a Gradio Web UI for local demos.
• Includes code for finetuning LLama base models with UniversalNER data.
• Evaluation code is available, with CrossNER and MIT datasets included due to licensing.

Maintenance & Community

The project is associated with authors from various institutions, indicating academic backing. Further community engagement channels are not explicitly listed in the README.

Licensing & Compatibility

The data and model checkpoints are intended and licensed for research use only. Usage is restricted by the license agreements of LLaMA, Vicuna, and ChatGPT.

Limitations & Caveats

Due to licensing restrictions, only a subset of NER datasets (CrossNER and MIT) are included for evaluation within the repository. The project's reliance on specific CUDA versions and LLM licenses may pose adoption challenges for commercial or broader research use.