A example-balanced mix of all five training formats: continuation, qa,
distillation, dialogue, take_comparison. Use meta.format to filter
or weight individual formats at training time.
Per-format counts and token totals are documented in
data/sft_mix_summary.json of the source repo.
example-balanced: each format down-sampled to the smallest format's… See the full description on the dataset page:
https://huggingface.co/datasets/abhayesian/ryan-greenblatt-sft-mix-examples-balanced.
