fariskishtah/entitylens-ai

EntityLens AI — BERT NER (bert-base-cased)

Fine-tuned bert-base-cased for English Named Entity Recognition on CoNLL-2003. Production model of the EntityLens AI project, selected from four architectures by validation entity-F1.

• Entities: PER, ORG, LOC, MISC (IOB2)
• Validation F1: 0.9537 · Test F1 (strict seqeval): 0.9178
• Per-class test F1: PER 0.963 · ORG 0.902 · LOC 0.931 · MISC 0.819

Usage

from transformers import pipeline
ner = pipeline("token-classification", model="fariskishtah/entitylens-ai",
               aggregation_strategy="simple")
ner("Barack Obama visited Microsoft in New York.")

Training

• 4 epochs, lr 5e-5, batch 16, max_len 256, first-subword label alignment.
• Trained on a Kaggle T4 GPU; strict entity-level seqeval evaluation.
• See the GitHub repo for the full pipeline, the three recurrent baselines (LSTM / BiLSTM / BiLSTM+CRF) and the CRF analysis.

Limitations

Trained on 1990s Reuters news; performance drops on other domains and on entity types outside PER/ORG/LOC/MISC. Confidence scores are raw softmax outputs, not calibrated probabilities.