Qsevent77/xlm-roberta-flash-implementation

Core implementation of Jina XLM-RoBERTa

This implementation is adapted from XLM-Roberta. In contrast to the original implementation, this model uses Rotary positional encodings and supports flash-attention 2.

Models that use this implementation

• jinaai/jina-embeddings-v3
• jinaai/jina-colbert-v2

Converting weights

Weights from an original XLMRoberta model can be converted using the convert_roberta_weights_to_flash.py script in the model repository.