I was exploring the original BERT model from 2018, which is mainly useful if you fine-tune a model on top of it for a specific task.
dslim/distilbert-NER by David S. Lim is a popular implementation of this, with around 20,000 downloads from Hugging Face every month.
I tried the demo from the README but it didn't quite work - it complained about an incompatibility with Numpy 2.0.
So I used uv run --with 'numpy<2.0' to run it in a temporary virtual environment. Here's a Bash one-liner that demonstrated the model:
uv run --with 'numpy<2.0' --with transformers python -c '
from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline
import json
model = AutoModelForTokenClassification.from_pretrained("dslim/distilbert-NER")
tokenizer = AutoTokenizer.from_pretrained("dslim/distilbert-NER")
nlp = pipeline("ner", model=model, tokenizer=tokenizer)
text = "This is an example sentence about Simon Willison who lives in Half Moon Bay"
print(json.dumps(nlp(text), indent=2, default=repr))'The first time you run this it will download 250MB to your ~/.cache/huggingface/hub/models--dslim--distilbert-NER folder.
Example output:
[
{
"entity": "B-PER",
"score": "0.9982101",
"index": 7,
"word": "Simon",
"start": 34,
"end": 39
},
{
"entity": "I-PER",
"score": "0.99835676",
"index": 8,
"word": "Willis",
"start": 40,
"end": 46
},
{
"entity": "I-PER",
"score": "0.9977602",
"index": 9,
"word": "##on",
"start": 46,
"end": 48
},
{
"entity": "B-LOC",
"score": "0.99432063",
"index": 13,
"word": "Half",
"start": 62,
"end": 66
},
{
"entity": "I-LOC",
"score": "0.99325883",
"index": 14,
"word": "Moon",
"start": 67,
"end": 71
},
{
"entity": "I-LOC",
"score": "0.9919292",
"index": 15,
"word": "Bay",
"start": 72,
"end": 75
}
]Created 2024-12-23T21:59:27-08:00 · Edit