---
pipeline_tag: fill-mask
license: mit
base_model: FacebookAI/roberta-large
library_name: kerasformers
tags:
- keras
- kerasformers
- roberta
- fill-mask
- text-encoder
- arxiv:1907.11692
- pytorch
- jax
- tf
---

## ***See [our collection](https://huggingface.co/collections/kerasformers/roberta-6a6e8d9b2f4c9253f4145f65) for all versions of RoBERTa.***

# Run RoBERTa with Keras 3: JAX, PyTorch, or TensorFlow

[![GitHub](https://img.shields.io/badge/GitHub-KerasFormers-black?logo=github)](https://github.com/IMvision12/KerasFormers) [![Docs](https://img.shields.io/badge/Docs-RoBERTa-blue)](https://imvision12.github.io/KerasFormers/roberta/) [![Collection](https://img.shields.io/badge/HF-RoBERTa%20collection-yellow)](https://huggingface.co/collections/kerasformers/roberta-6a6e8d9b2f4c9253f4145f65)

# kerasformers/roberta_large

Paper: [RoBERTa: A Robustly Optimized BERT Pretraining Approach (arXiv:1907.11692)](https://arxiv.org/abs/1907.11692) · [HF Papers](https://huggingface.co/papers/1907.11692)

RoBERTa is a robustly optimized BERT encoder: more data/steps, no NSP, dynamic masking, byte-level BPE (mask token `<mask>`), and padding-offset position ids.

For more details on the model, please go to the upstream [model card](https://huggingface.co/FacebookAI/roberta-large).

Pure-**Keras 3** conversion of [`FacebookAI/roberta-large`](https://huggingface.co/FacebookAI/roberta-large) for [kerasformers](https://github.com/IMvision12/KerasFormers). One implementation runs unmodified on **TensorFlow / Torch / JAX**.

This is a **fill-mask / encoder** checkpoint (`RobertaMaskedLM`, large). Task heads load via `hf:` fine-tunes.

## ✨ Quick start (fill-mask)

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

from kerasformers.models.roberta import RobertaMaskedLM, RobertaTokenizer

mlm = RobertaMaskedLM.from_weights("kerasformers/roberta_large")
tokenizer = RobertaTokenizer.from_weights("kerasformers/roberta_large")

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 RoBERTa variant the same way with `from_weights("kerasformers/<variant>")`:

| Variant | Hub |
|---|---|
| `roberta_base` | [`kerasformers/roberta_base`](https://huggingface.co/kerasformers/roberta_base) |
| `roberta_large` | [`kerasformers/roberta_large`](https://huggingface.co/kerasformers/roberta_large) |

## Tips

- Set `KERAS_BACKEND` **before** importing Keras / kerasformers.
- Prefer `RobertaTokenizer.from_weights(...)` so BPE vocab matches.
- Use `<mask>` (not `[MASK]`).
- See [RoBERTa docs](https://imvision12.github.io/KerasFormers/roberta/) and [Loading Weights](https://imvision12.github.io/KerasFormers/loading_weights/).
- Community / upstream safetensors still work via the `hf:` prefix, e.g. `RobertaMaskedLM.from_weights("hf:FacebookAI/roberta-large")`.

## Special Thanks

A huge thank you to the Facebook AI RoBERTa authors for creating and releasing these models.

License: MIT.