README.md

---
library_name: transformers
license: apache-2.0
pipeline_tag: feature-extraction
---

# Overview

This repository contains an encoder model, part of the research presented in the paper *Should We Still Pretrain Encoders with Masked Language Modeling?* (Gisserot-Boukhlef et al.).

*   **Paper:** [Should We Still Pretrain Encoders with Masked Language Modeling?](https://huggingface.co/papers/2507.00994)
*   **Blog post:** [Link](https://huggingface.co/blog/Nicolas-BZRD/encoders-should-not-be-only-pre-trained-with-mlm)
*   **Project page:** [https://hf.co/MLMvsCLM](https://hf.co/MLMvsCLM)
*   **Code:** [https://github.com/Nicolas-BZRD/EuroBERT](https://github.com/Nicolas-BZRD/EuroBERT)

## Model Naming

Model identifiers follow a consistent format that encodes key training details:

* **Single-stage models**:
`[model size]-[objective]-[number of steps]`.
Example: `610m-clm-42k` denotes a 610M-parameter model trained with CLM for 42,000 steps.
* **Two-stage models**:
`[model size]-[objective #1]-[steps #1]-[objective #2]-[total steps]`.
Example: `610m-clm-10k-mlm40-42k` indicates a 610M model trained first with CLM for 10k steps, then continued with MLM (40% masking ratio) for 32k more steps, totaling 42k steps.
* **Continued pretraining from decayed checkpoints**:
These use the dec prefix on the first training stage.
Example: `610m-clm-dec42k-mlm40-64k refers` to a 610M model pretrained with CLM for 42k steps (with weight decay), then further trained with MLM (40% masking) for 22k additional steps, totaling 64k.
* **Intermediate checkpoints**:
To refer to a specific training step before the final checkpoint, append the step number at the end.
Example: `610m-mlm40-42k-1000` corresponds to step 1,000 during the MLM training phase of a 610M model trained for 42k steps.

## Usage

You can use this model for feature extraction with the Hugging Face `transformers` library.

```python
from transformers import AutoTokenizer, AutoModel
import torch

# Replace with the actual model ID if different, e.g., "AhmedAliHassan/MLMvsCLM-Biphasic-210M"
# This placeholder assumes the current repository is the model you want to load.
model_name = "<YOUR_MODEL_ID_HERE>" 

# Load the tokenizer and model, ensuring trust_remote_code for custom architectures
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, trust_remote_code=True)

text = "This is an example sentence to extract features from."

inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

# The last hidden state contains the token embeddings (features)
last_hidden_state = outputs.last_hidden_state
print(f"Shape of last hidden state: {last_hidden_state.shape}")

# For sentence-level embeddings, common approaches include:
# 1. Averaging the token embeddings (excluding special tokens)
# 2. Using the embedding of the [CLS] token (if applicable for the model's architecture)
# Example: Mean pooling (simple average over non-padding tokens)
attention_mask = inputs["attention_mask"]
input_mask_expanded = attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float()
sum_embeddings = torch.sum(last_hidden_state * input_mask_expanded, 1)
sum_mask = torch.clamp(input_mask_expanded.sum(1), min=1e-9)
mean_pooled_embedding = sum_embeddings / sum_mask
print(f"Shape of mean pooled embedding: {mean_pooled_embedding.shape}")
```

## Citation

If you found this model useful, please consider citing our paper:

```bibtex
@misc{gisserotboukhlef2025pretrainencodersmaskedlanguage,
      title={Should We Still Pretrain Encoders with Masked Language Modeling?}, 
      author={Hippolyte Gisserot-Boukhlef and Nicolas Boizard and Manuel Faysse and Duarte M. Alves and Emmanuel Malherbe and André F. T. Martins and Céline Hudelot and Pierre Colombo},
      year={2025},
      eprint={2507.00994},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.00994}, 
}
```