---

name: flair description: "Use Flair for NLP annotation, embeddings, text/sequence model prediction, corpus loading, training/fine-tuning, HunFlair biomedical NER/linking, and production troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Flair

Use this skill when a task names the public flair Python package or asks for Flair-specific NLP workflows: Sentence annotations, pretrained taggers/classifiers, token/document embeddings, custom corpora, model training/fine-tuning, HunFlair/HunFlair2 biomedical NER/linking, or Flair cache/device troubleshooting.

Install and baseline check

Flair 0.15.1 requires Python 3.9+ and PyTorch. A normal package-user starting point is:

pip install flair
python - <<'PY'
import flair
print(flair.__version__)
print(flair.device)
print(flair.cache_root)
PY

Set runtime policy before importing flair:

export FLAIR_DEVICE=cpu                 # verified baseline for this skill
export FLAIR_CACHE_ROOT="./flair-cache" # optional deliberate cache location

Named pretrained models, embeddings, datasets, and biomedical dictionaries may download public resources if they are not already cached. CUDA, multi-GPU, ONNX/provider runtimes, SciSpaCy, pyab3p, OCR/layout inputs, and large model downloads are optional paths unless the active environment verifies them.

For a safe package diagnostic that avoids model/data downloads, run scripts/collect_env.py with --json or --check-imports.

Route by task

• Annotations, tokenization, prediction outputs, regex tagging, serialization, or visualization: use tagging-and-annotations for Sentence, Token, Span, Relation, Label, DataPair, Classifier.load(...), label layers, tokenizers, splitters, RegexpTagger, Sentence.to_dict(), and NER HTML rendering.
• Embedding choice, vector extraction, language-model embeddings, cache/device behavior, long-sentence/context, or optional ONNX/JIT optimization: use embeddings-and-optimization for Transformer*Embeddings, WordEmbeddings, FlairEmbeddings, StackedEmbeddings, DocumentPool/RNN/CNNEmbeddings, LanguageModel, and tensor shape checks.
• Custom corpora, label dictionaries, model construction, training/fine-tuning, checkpoints, TARS, multitask, relation/span training, or optional multi-GPU: use training-and-datasets for ColumnCorpus, ClassificationCorpus, CSVClassificationCorpus, JSONL, CoNLL-U, MultiCorpus, ModelTrainer.train, ModelTrainer.fine_tune, and the bundled NER helper.
• Biomedical NER/linking: use biomedical-nlp for HunFlair/HunFlair2, EntityMentionLinker, biomedical dictionaries, SciSpaCy tokenization, pyab3p abbreviation behavior, biomedical corpus offsets, and NER-vs-linking layer separation.

Common starting points

• For first-time Flair prediction: create Sentence(...), load a model with Classifier.load(model_id) only when downloads/cache are allowed, call predict, then read the exact layer with sentence.get_labels("layer") or sentence.get_spans("layer").
• For a no-download annotation check: run sub-skills/tagging-and-annotations/scripts/annotation_smoke.py --json.
• For a no-download embedding check: run sub-skills/embeddings-and-optimization/scripts/embedding_smoke.py --json.
• For corpus/training planning: run sub-skills/training-and-datasets/scripts/fine_tune_ner.py --help, --list-datasets, or --dry-run before training.
• For biomedical linking without downloads: run sub-skills/biomedical-nlp/scripts/biomedical_smoke.py --run-local-linker --json.

Shared references

• references/api-overview.md: verified package identity, key APIs/signatures, import caveats, and cross-cutting data/model objects.
• references/model-selection.md: task-to-model and embedding selection guide, including pretrained model IDs and when to use Classifier.load.
• references/troubleshooting.md: install/import, cache, device, download, optional dependency, and cross-sub-skill failure recovery.
• references/repo-provenance.md: source snapshot, package version, evidence paths, and staleness checks.

Boundaries

This skill is a runtime package-use guide, not a maintainer release or CI guide. It does not cover website generation, release publishing, broad lint/type-check stacks, private datasets, long benchmark training, or quality claims for unrun optional backends. If a current checkout or installed Flair version differs materially from the provenance snapshot, refresh the skill before relying on exact API details.