README.md

---
pipeline_tag: fill-mask
license: apache-2.0
base_model: google/electra-base-generator
library_name: zeromodels
tags:
- keras
- zeromodels
- electra
- generator
- text-encoder
- fill-mask
- arxiv:2003.10555
- pytorch
- jax
- tf
---

## ***See [our collection](https://huggingface.co/collections/zeromodels/electra-6a8eadf9dc472c12a679ebba) for all versions of ELECTRA.***

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

[![GitHub](https://img.shields.io/badge/GitHub-ZeroModels-black?logo=github)](https://github.com/IMvision12/ZeroModels) [![Docs](https://img.shields.io/badge/Docs-ELECTRA-blue)](https://imvision12.github.io/ZeroModels/electra/) [![Collection](https://img.shields.io/badge/HF-ELECTRA%20collection-yellow)](https://huggingface.co/collections/zeromodels/electra-6a8eadf9dc472c12a679ebba)

# zeromodels/electra_base_generator

Paper: [ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators (arXiv:2003.10555)](https://arxiv.org/abs/2003.10555) · [HF Papers](https://huggingface.co/papers/2003.10555)

ELECTRA is Google's BERT-style bidirectional text encoder, pre-trained as a replaced-token **discriminator** (with a smaller **generator** producing the corrupted tokens). This repo is the **masked-LM (fill-mask)** checkpoint. WordPiece tokenizer; mask token `[MASK]`.

For more details on the model, please go to the upstream [model card](https://huggingface.co/google/electra-base-generator).

Pure-**Keras 3** conversion of [`google/electra-base-generator`](https://huggingface.co/google/electra-base-generator) for [zeromodels](https://github.com/IMvision12/ZeroModels). One implementation runs unmodified on **TensorFlow / Torch / JAX**.

## ✨ Quick start (masked-LM (fill-mask))

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

from zeromodels.models.electra import ElectraMaskedLM, ElectraTokenizer

mlm = ElectraMaskedLM.from_weights("zeromodels/electra_base_generator")
tokenizer = ElectraTokenizer.from_weights("zeromodels/electra_base_generator")

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 ELECTRA variant the same way with `from_weights("zeromodels/<variant>")`:

| Size | Discriminator (encoder / downstream) | Generator (masked-LM) |
|---|---|---|
| small | [`zeromodels/electra_small_discriminator`](https://huggingface.co/zeromodels/electra_small_discriminator) | [`zeromodels/electra_small_generator`](https://huggingface.co/zeromodels/electra_small_generator) |
| base | [`zeromodels/electra_base_discriminator`](https://huggingface.co/zeromodels/electra_base_discriminator) | [`zeromodels/electra_base_generator`](https://huggingface.co/zeromodels/electra_base_generator) |
| large | [`zeromodels/electra_large_discriminator`](https://huggingface.co/zeromodels/electra_large_discriminator) | [`zeromodels/electra_large_generator`](https://huggingface.co/zeromodels/electra_large_generator) |

## Available classes

Load any of these from this repo with `from_weights("zeromodels/electra_base_generator")` (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 |
|---|---|
| `ElectraMaskedLM` | Masked language modeling (fill-mask) |

```python
from zeromodels.models.electra import ElectraMaskedLM
model = ElectraMaskedLM.from_weights("zeromodels/electra_base_generator")
```

## Tips

- Set `KERAS_BACKEND` **before** importing Keras / zeromodels.
- Prefer `ElectraTokenizer.from_weights(...)` so WordPiece tokenization matches.
- Downstream tasks (classification / QA / NER) use the **discriminator** repos; the **generator** repos are the masked-LM.
- See [ELECTRA docs](https://imvision12.github.io/ZeroModels/electra/) and [Loading Weights](https://imvision12.github.io/ZeroModels/loading_weights/).
- Community / upstream safetensors still work via the `hf:` prefix, e.g. `ElectraModel.from_weights("hf:google/electra-base-generator")`.

## Special Thanks

A huge thank you to the Google ELECTRA authors for creating and releasing these models.

License: Apache 2.0.