huggingface-transformers

Hugging Face Transformers AI Skill Guide

Overview & Engine Architecture

Transformers provides pretrained model configs, tokenizers, and Auto* loaders plus high-level pipeline helpers. Tokenizers map text to tensors; models run on PyTorch/TensorFlow/Flax backends. Agents pin model revisions, respect max sequence lengths, separate train/eval modes, and treat Hub downloads as supply-chain inputs (revision hashes, not floating latest).

Tokenizer -> input_ids / attention_mask
      -> AutoModel* (forward)
          -> logits / generated tokens
          -> decode

When to use this skill

• NLP classification, NER, summarization, generation
• Vision/audio models exposed via Transformers APIs
• Fine-tuning with Trainer or custom @pytorch loops

Operational directives

1. Pin revision (commit hash) for production model loads.
2. Use pipeline for prototypes; switch to explicit tokenizer+model for control.
3. Truncate/pad consistently with the model's max length.
4. Call model.eval() and torch.inference_mode() for serving paths.
5. Respect model licenses and data privacy before uploading to the Hub.

Pipeline + explicit inference

from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification
import torch

clf = pipeline("text-classification", model="distilbert-base-uncased-finetuned-sst-2-english", revision="main")
print(clf("This deployment looks solid."))

tok = AutoTokenizer.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english")
model = AutoModelForSequenceClassification.from_pretrained(
    "distilbert-base-uncased-finetuned-sst-2-english",
    revision="main",
)
model.eval()
batch = tok(["ship it", "needs work"], return_tensors="pt", padding=True, truncation=True)
with torch.inference_mode():
    logits = model(**batch).logits
    print(logits.softmax(-1))

Fine-tune sketch

from transformers import Trainer, TrainingArguments

args = TrainingArguments(
    output_dir="out/sentiment",
    per_device_train_batch_size=16,
    num_train_epochs=2,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
)
trainer = Trainer(model=model, args=args, train_dataset=train_ds, eval_dataset=val_ds, tokenizer=tok)
trainer.train()