bert-base-cased

Predicts masked tokens in text using bidirectional transformer attention, where the model attends to both left and right context simultaneously. Implements the MLM (Masked Language Modeling) objective trained on BookCorpus and Wikipedia, enabling it to infer missing words based on surrounding context. Uses 12 transformer layers with 768 hidden dimensions and 12 attention heads, processing input through WordPiece tokenization (30,522 vocabulary tokens) and returning logits across the full vocabulary for each masked position.

Extracts learned token representations from the model's hidden layers, producing dense vector embeddings (768-dimensional) for each input token. The model learns these embeddings through unsupervised pretraining on masked language modeling and next-sentence-prediction objectives, capturing semantic and syntactic relationships. Embeddings can be extracted from any of the 12 transformer layers, with later layers capturing more task-specific information and earlier layers capturing more syntactic patterns.

Predicts whether two text segments are consecutive sentences in the original document using a binary classification head trained during pretraining. The model encodes both segments with a [SEP] token separator and [CLS] token prefix, then uses the [CLS] token's final hidden state (passed through a dense layer) to output a binary logit. This was trained on 50% positive pairs (consecutive sentences) and 50% negative pairs (random sentences), enabling the model to learn document-level coherence patterns.

Supports loading and inference across PyTorch, TensorFlow, and JAX/Flax frameworks through a unified HuggingFace Transformers API, with automatic weight conversion and framework-specific optimizations. The model weights are stored in SafeTensors format (binary serialization with built-in integrity checks) and can be loaded into any framework without manual conversion. Transformers library handles tokenization, batching, and framework-specific device placement (CPU/GPU/TPU) transparently.

Tokenizes input text into subword units using WordPiece algorithm with a case-sensitive 30,522-token vocabulary, preserving case distinctions (e.g., 'Apple' vs 'apple' are different tokens). The tokenizer uses greedy longest-match-first algorithm to split unknown words into subword units prefixed with '##' (e.g., 'unbelievable' → ['un', '##believ', '##able']). Special tokens include [CLS] (sequence start), [SEP] (segment separator), [MASK] (masked position), [UNK] (unknown), [PAD] (padding).

Enables transfer learning by freezing or unfreezing pretrained transformer weights and adding task-specific classification heads (linear layers) on top of BERT's output. The model can be fine-tuned end-to-end (all layers trainable) or with selective unfreezing (e.g., only top 2-4 layers + classification head). Supports standard supervised learning with cross-entropy loss, with learning rates typically 1e-5 to 5e-5 to avoid catastrophic forgetting of pretrained knowledge.

Exposes attention weights from all 12 transformer layers and 12 attention heads, enabling visualization of which input tokens the model attends to when predicting each output token. Attention weights are returned as tensors (shape: batch_size × num_heads × sequence_length × sequence_length) and can be aggregated across heads or layers to identify important token relationships. This enables analysis of what linguistic patterns the model learns (e.g., attention to pronouns for coreference, attention to punctuation for syntax).

Processes multiple input sequences in parallel with automatic dynamic padding (padding to longest sequence in batch rather than fixed length), reducing computation on short sequences. The tokenizer returns attention_mask tensors indicating which positions are padding, allowing the model to ignore padded positions in attention computation. Batching is handled transparently by the Transformers library, with configurable batch sizes and automatic device placement (CPU/GPU).

Leverages pretrained representations learned from 3.3B token corpus to perform zero-shot or few-shot inference on tasks not explicitly trained on, by using embeddings or attention patterns as features for downstream classifiers. The model's learned linguistic knowledge (syntax, semantics, named entities) transfers to new tasks without fine-tuning, though performance is typically lower than fine-tuned models. Common approaches include using [CLS] embeddings as document features or using attention patterns for task-specific signals.

While trained exclusively on English, BERT-base-cased can perform cross-lingual transfer through shared subword vocabulary with other languages, where multilingual BERT variants share the same WordPiece tokenizer across 104 languages. For English-only BERT, this means the 30K vocabulary contains some non-English tokens learned incidentally during pretraining, enabling limited transfer to similar languages (German, Dutch, French) through shared vocabulary overlap. This is not true multilingual support but rather vocabulary-based transfer.