๐Ÿ“Š Monitoring#

This guide gives you a brief introduction to Argilla Metrics. Argilla Metrics enable you to perform fine-grained analyses of your models and training datasets. Argilla Metrics are inspired by a a number of seminal works such as Explainaboard.

The main goal is to make it easier to build more robust models and training data, going beyond single-number metrics (e.g., F1).

This guide gives a brief overview of currently supported metrics. For the full API documentation see the Python API reference

This feature is experimental, you can expect some changes in the Python API. Please report on Github any issue you encounter.

Install dependencies#

Verify you have already installed Jupyter Widgets in order to properly visualize the plots. See https://ipywidgets.readthedocs.io/en/latest/user_install.html

For running this guide you need to install the following dependencies:

[ ]:
%pip install datasets spacy plotly -qqq

and the spacy model:

[ ]:
!python -m spacy download en_core_web_sm

1. NER prediction metrics#

Load dataset and model#

Weโ€™ll be using spaCy for this guide, but all the metrics weโ€™ll see are computed for any other framework (Flair, Stanza, Hugging Face, etc.). As an example will use the WNUT17 NER dataset.

[ ]:
import argilla as rg
import spacy
from datasets import load_dataset

nlp = spacy.load("en_core_web_sm")
dataset = load_dataset("wnut_17", split="train")

Log records in dataset#

Letโ€™s log spaCy predictions using the built-in rg.monitor method:

[ ]:
nlp = rg.monitor(nlp, dataset="spacy_sm_wnut17")

def predict(record):
    doc = nlp(" ".join(record["tokens"]))
    return {"predicted": True}


Explore pipeline metrics#

from argilla.metrics.token_classification import token_length


from argilla.metrics.token_classification import token_capitalness


from argilla.metrics.token_classification import token_frequency

token_frequency(name="spacy_sm_wnut17", tokens=50).visualize()

from argilla.metrics.token_classification import top_k_mentions

top_k_mentions(name="spacy_sm_wnut17", k=5000, threshold=2).visualize()

from argilla.metrics.token_classification import entity_labels


from argilla.metrics.token_classification import entity_density


from argilla.metrics.token_classification import entity_capitalness


from argilla.metrics.token_classification import mention_length


2. NER training Metrics#

Analyze tags#

Letโ€™s analyze the conll2002 dataset at the tag level.

[ ]:
dataset = load_dataset("conll2002", "es", split="train[0:5000]")

def parse_entities(record):
    entities = []
    counter = 0
    for i in range(len(record["ner_tags"])):
        entity = (
            counter + len(record["tokens"][i]),
        counter += len(record["tokens"][i]) + 1
    return entities

records = [
        text=" ".join(example["tokens"]),
    for example in dataset

[ ]:
rg.log(records, "conll2002_es")

from argilla.metrics.token_classification import top_k_mentions
from argilla.metrics.token_classification.metrics import Annotations

    name="conll2002_es", k=30, threshold=4, compute_for=Annotations

From the above we see we can quickly detect an annotation issue: double quotes " are most of the time tagged as O (no entity) but in some cases (~60 examples) are tagged as beginning of entities like ORG or MISC, which is likely a hand-labelling error, including the quotes inside the entity span.

from argilla.metrics.token_classification import *

entity_density(name="conll2002_es", compute_for=Annotations).visualize()

3. TextClassification metrics#

[ ]:
from datasets import load_dataset
from transformers import pipeline

import argilla as rg

sst2 = load_dataset("glue", "sst2", split="validation")
labels = sst2.features["label"].names
nlp = pipeline("sentiment-analysis")

records = [
            (pred["label"].lower(), pred["score"]) for pred in nlp(record["sentence"])
    for record in sst2

[ ]:
rg.log(records, name="sst2")

from argilla.metrics.text_classification import f1


# now compute metrics for negation ( -> negative precision and positive recall go down)
f1(name="sst2", query="n't OR not").visualize()