---
library_name: keras-hub
---
### Model Overview
## Model Summary

Multilingual E5 is a family of multilingual text embedding models from [intfloat ](https://huggingface.co/intfloat), based on the XLM-RoBERTa architecture. These models generate fixed-size sentence embeddings suitable for semantic search, retrieval, clustering, and text classification across 100+ languages.

The E5 models were introduced in ["Multilingual E5 Text Embeddings: A Technical Report"](https://arxiv.org/abs/2402.05672)  by Wang et al. The models are trained in two stages: (1) contrastive pre-training on ~1 billion multilingual text pairs, and (2) fine-tuning on labeled datasets. Input text should be prefixed with "query: " or "passage: " for optimal performance.

## Key Features:
- Supports 100+ languages via XLM-RoBERTa backbone
- Trained on ~1 billion multilingual text pairs
- MIT licensed
- Works out of the box for semantic search, retrieval, clustering, and classification
- Model Family: Multilingual E5
- Architecture: XLM-RoBERTa (bidirectional Transformer encoder)
- Languages: 100+
- Task Type: Text Embedding / Sentence Similarity / Retrieval
- Max Sequence Length: 512 tokens

## Model Details

* [Multilingual E5 Quickstart Notebook](coming soon..)
* [Multilingual E5 API Documentation](https://keras.io/keras_hub/api/models/xlm_roberta/)
* [Multilingual E5 Model Card](https://huggingface.co/intfloat/multilingual-e5-small)
* [Multilingual E5 Technical Paper](https://arxiv.org/abs/2402.05672)
* [KerasHub Beginner Guide](https://keras.io/guides/keras_hub/getting_started/)
* [KerasHub Model Publishing Guide](https://keras.io/guides/keras_hub/upload/)

## Installation

Keras and KerasHub can be installed with:

```
pip install -U -q keras-hub
pip install -U -q keras

```

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the [Keras Getting Started](https://keras.io/getting_started/) page.

## Presets

## Preset Table

| Preset | Architecture | Pooling | Normalize | Languages | Description |
|---|---|---|---|---|---|
| `multilingual_e5_small` | XLM-RoBERTa | Mean | L2 | 100+ | 12-layer Multilingual E5 small model. Produces 384-dim embeddings. |
| `multilingual_e5_base` | XLM-RoBERTa | Mean | L2 | 100+ | 12-layer Multilingual E5 base model. Produces 768-dim embeddings. |
| `multilingual_e5_large` | XLM-RoBERTa | Mean | L2 | 100+ | 24-layer Multilingual E5 large model. Produces 1024-dim embeddings. |

## Example Usage
```
import keras_hub

# Load the text embedder with preprocessing.
embedder = keras_hub.models.TextEmbedder.from_preset(
    "multilingual_e5_small"
)

# Encode queries
query = "query: Which planet is known as the Red Planet?"
q_emb = embedder.encode_text(query)

# Encode documents/passages
documents = [
    "passage: Mars is often referred to as the Red Planet.",
    "passage: Venus is often called Earth's twin.",
]
d_embs = embedder.encode_text(documents)
```

```
# Load just the backbone for custom architectures.
backbone = keras_hub.models.XLMRobertaBackbone.from_preset(
    "multilingual_e5_small",
)
```

## Example Usage with Hugging Face URI

```
import keras_hub

# Load the text embedder with preprocessing.
embedder = keras_hub.models.TextEmbedder.from_preset(
    "hf://keras/multilingual_e5_small"
)

# Encode queries
query = "query: Which planet is known as the Red Planet?"
q_emb = embedder.encode_text(query)

# Encode documents/passages
documents = [
    "passage: Mars is often referred to as the Red Planet.",
    "passage: Venus is often called Earth's twin.",
]
d_embs = embedder.encode_text(documents)
```

```
# Load just the backbone for custom architectures.
backbone = keras_hub.models.XLMRobertaBackbone.from_preset(
    "hf://keras/multilingual_e5_small",
)
```