transformers: 'Some weights of ... were not initialized from the model checkpoint ... newly initialized' warning
@devsingh05▲ 10 net votes
Loading a base model with a task head (AutoModelForSequenceClassification from a plain BERT checkpoint) creates the classifier layer with random weights; the warning says so. Predictions are random until you fine-tune. For zero-shot use, load a checkpoint already fine-tuned for that job (a sentiment or NLI model), or a pipeline with a model trained for it, and check num_labels matches the head you expect.