How did you create (title, paragraph) pairs from c4?
#45
by
itayair
- opened
In addition, did you filter the C4 data?
To create contrastive pairs, please refer to the discussion at https://huggingface.co/intfloat/multilingual-e5-large/discussions/37#664b1fe87a1ed3e001471b2f
And yes, we filter the mC4 data using the consistency-based filtering approach in Text Embeddings by Weakly-Supervised Contrastive Pre-training
One more thing, what does the page_content you take? (The web pages might be much longer than 512)
In that case, texts will be truncated to fit the model's maximum support length.