Update DFlash to 1M-verifier continuation step 15000
#2
by baranowskiadam - opened
Weight-only continuation checkpoint trained against the 1M YaRN-128 polishing verifier. The DFlash preserves its trained default theta-500000 RoPE; max_position_embeddings is extended to 1048576 for compatibility with poolside/Laguna-S-2.1-polishing.
joerowell changed pull request status to merged