Run ModernBERT with Keras 3: JAX, PyTorch, or TensorFlow

kerasformers/modernbert_base

Paper: Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder (arXiv:2412.13663) · HF Papers

ModernBERT is Answer.AI / LightOn's modernized bidirectional transformer text encoder: rotary position embeddings, attention that alternates between a global (full) layer and local sliding-window layers, GeGLU feed-forwards, and pre-LayerNorm, with an 8192-token context. Byte-level BPE tokenizer; mask token [MASK]. No token-type ids.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of answerdotai/ModernBERT-base for kerasformers. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is a fill-mask / encoder checkpoint (ModernBertMaskedLM, modernbert_base). Task heads (sequence / token classify, QA, multiple choice) load via hf: fine-tunes.

✨ Quick start (fill-mask)

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from kerasformers.models.modernbert import ModernBertMaskedLM, ModernBertTokenizer

mlm = ModernBertMaskedLM.from_weights("kerasformers/modernbert_base")
tokenizer = ModernBertTokenizer.from_weights("kerasformers/modernbert_base")

inputs = tokenizer("The capital of France is [MASK].")
logits = mlm(inputs)  # (1, L, vocab_size)
mask = int((inputs["input_ids"][0] == tokenizer.mask_token_id).argmax())
print(tokenizer.decode([int(logits[0, mask].argmax())]))

Load any ModernBERT variant the same way with from_weights("kerasformers/<variant>"):

Variant	Hub	layers	embed_dim
`modernbert_base`	`kerasformers/modernbert_base`	22	768
`modernbert_large`	`kerasformers/modernbert_large`	28	1024

Available classes

Load any of these from this repo with from_weights("kerasformers/modernbert_base") (or on the fly via the hf: prefix). The pretrained backbone is shared; task heads not stored in this checkpoint start randomly initialized, ready for fine-tuning (or load a hf: fine-tune).

Class	Task
`ModernBertModel`	Encoder backbone
`ModernBertMaskedLM`	Masked language modeling (fill-mask)
`ModernBertSequenceClassify`	Sequence classification
`ModernBertTokenClassify`	Token classification (NER / POS)
`ModernBertQnA`	Extractive question answering
`ModernBertMultipleChoice`	Multiple choice

from kerasformers.models.modernbert import ModernBertSequenceClassify
model = ModernBertSequenceClassify.from_weights("kerasformers/modernbert_base")

Tips

Set KERAS_BACKEND before importing Keras / kerasformers.
Prefer ModernBertTokenizer.from_weights(...) so tokenization matches.
Use [MASK] (not <mask>).
ModernBERT has no token-type ids; the tokenizer emits only input_ids / attention_mask.
See ModernBERT docs and Loading Weights.
Community / upstream safetensors still work via the hf: prefix, e.g. ModernBertMaskedLM.from_weights("hf:answerdotai/ModernBERT-base").