No way to train speech interface!

Hello, the speech datasets' links for json and features all link back to the xllm repo. The speech encoder link also links back to the main page of xllm. Where do we find the blip2 checkpoints? Do we use their Q former only?