Flair NLP Cheat Sheet

Flair cheat sheet

Flair is a framework for NLP with a famously simple API: wrap text in a Sentence, load a pretrained Classifier, call predict, read the labels. It shines at NER, POS & text classification, and at stacking embeddings (classic word vectors + its own contextual Flair embeddings + Transformers) to train strong custom models. Built on PyTorch, models hosted on the 🌐 Hub. This sheet targets Flair 0.15 (Python 3.9+).

Sentence & predict tagging tasks embeddings training interop gotcha ★ most common

Verified 2026-08-24 against the official docs at flairnlp.github.io and the flairNLP/flair repo (0.15.1). PyTorch-based; pretrained models on the Hugging Face Hub. Compare with spaCy (production) & Stanza (multilingual accuracy).

Outline

Inference is three lines: Sentence → Classifier.load → predict. The depth is in embeddings & training your own tagger.

1Install, Sentence, predict0.15

pip install flair★
Pulls in PyTorch + Transformers. Python 3.9+.
from flair.data import Sentence from flair.nn import Classifier★
Sentence is the core data object (it tokenizes for you); Classifier is the universal model loader.
sent = Sentence("George Washington went to Washington.") tagger = Classifier.load("ner") tagger.predict(sent)★
The whole inference flow. predict annotates the sentence in place — it returns None.
tagger.predict([sent1, sent2], mini_batch_size=32)
Pass a list to batch. Load the model once and reuse it.

2Reading labels & spansget results out

for ent in sent.get_spans("ner"): print(ent.text, ent.tag, ent.score)★
NER results are spans — multi-token entities with .text, .tag (label), and .score (confidence).
sent.labels · sent.get_labels("ner")★
All predicted labels on the sentence. For sentence-level tasks (sentiment), the label is on the sentence itself.
print(sent) · sent.to_dict()
Printing shows the tagged text; to_dict() for a serializable structure.
for token in sent: token.text, token.get_label("pos").value
Iterate tokens for per-token tags (POS). Each token also carries offsets & embeddings.

6Corpus & label dictyour data

from flair.datasets import ColumnCorpus corpus = ColumnCorpus("data/", {0: "text", 1: "ner"}, train_file="train.txt", dev_file=..., test_file=...)★
Load CoNLL-style column data (token → tag per line). Built-in corpora exist too (flair.datasets.CONLL_03).
label_dict = corpus.make_label_dictionary(label_type="ner")★
Build the set of target labels from the corpus — needed to construct the model.
corpus.train · corpus.dev · corpus.test · corpus.obtain_statistics()
Splits are ready-made; inspect class balance before training.

7Train a modelSequenceTagger / ModelTrainer

from flair.models import SequenceTagger tagger = SequenceTagger(hidden_size=256, embeddings=stacked, tag_dictionary=label_dict, tag_type="ner")★
A BiLSTM-CRF tagger over your embeddings. TextClassifier for document classification.
from flair.trainers import ModelTrainer trainer = ModelTrainer(tagger, corpus) trainer.train("out/", max_epochs=10)★
Trains, evaluates on dev/test, and saves the best model to out/best-model.pt.
trainer.fine_tune("out/", learning_rate=5e-6) # for transformer embeddings★
Use fine_tune, not train, with a Transformer embedding — it sets the right (low) LR & schedule for fine-tuning.
model = SequenceTagger.load("out/best-model.pt")
Reload your trained model and predict like any other.

!Common gotchasread before shipping

predict/embed mutate in place
They return None and annotate the Sentence object — read results off sent, don't expect a return value.
NER = spans, sentiment = sentence label
Use get_spans("ner") for entities but sent.labels for document-level tasks — different accessors.
fine_tune() for transformers
Calling train() with a Transformer embedding uses the wrong schedule. fine_tune() is the intended path.
embeddings stay in memory
Stacking large transformer + Flair embeddings is memory-heavy. Embed in mini-batches and reuse one loaded model.

Worth memorizing

Sentence → Classifier.load → predictthree lines; predict annotates in place

NER results are spansget_spans("ner") → .text / .tag / .score

load models by task or HF id"ner", "sentiment", "pos" or any hub model

embed() mutates the sentencevectors land on token.embedding / sent.embedding

StackedEmbeddings is the recipeGloVe + forward/backward Flair embeddings

Document vs Word embeddingsone vector per sentence vs per token

ColumnCorpus + make_label_dictionaryCoNLL columns → labels to train

fine_tune for transformers, train otherwisecorrect LR/schedule

Flair cheat sheet

Outline

Predict

Tasks

Embeddings

Train

Predict

Tasks

Embeddings

Train Your Own

Worth memorizing