Iteration focus: describe the CONTENT, not the SOURCE. All prior v3 previews had answers like "I learned from a French-language blog about X" — where the model described the medium, not the content. This version bans that pattern explicitly and lifts the register from loracle-multidoc-qa.

	Key changes vs earlier v3 preview… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-pretrain-qa-v3b-preview1k.