IC-1465InstructUIE-11B achieves 81.16% average F1 on 20 in-domain NER datasets and 49.4% on out-of-domain evaluation

Wenxuan Zhou, Sheng Zhang, Yu Gu, Muhao Chen, Hoifung Poon

SourceUniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition

InstructUIE, a multi-task instruction-tuned system based on FLAN-T5 11B that uses supervised NER examples, is evaluated in both in-domain and out-of-domain settings. In the in-domain setting (Table 2, 20 datasets from Wang et al. 2023a), it achieves 81.16% average F1. In the out-of-domain setting (Table 3, excluding CrossNER and MIT datasets from training), it achieves 49.4% average F1. The paper notes a critical bug in InstructUIE's original evaluation script that inflates F1 for passages without entities, and re-evaluated using the released checkpoint.

Evidence
correlational
Key metric
81.16% avg F1 (in-domain, 20 datasets); 49.4% avg F1 (out-of-domain, 7 domains)
Caveat
The original InstructUIE evaluation script contained a bug where 'none' was added as a placeholder entity for entity-free passages, inflating F1; the paper re-evaluated using the released checkpoint to correct this.
Model
InstructUIE
Related findings
IC-1463, IC-1464
Extraction
automatic-extraction