Bi-encoder model trained on paraphrases and new generated questions based on subset from datasets:
Context length was cut to 1024 tokens.
Files info