permissive-post-train 's Collections

Front-loading study

Four web datasets pretrained with and without MixtureVitae's instruction and reasoning subset, then extended to 16K and given Tulu3 SFT and DPO.