Smolcluster GRPO favors staged curricula

This reads less like a model release and more like a useful lesson in reward design: for tiny summarizers, the order of objectives matters more than stacking every signal at once.

• –Staged training beat joint training across both base models, which suggests the length constraint is doing real optimization work rather than acting as a cosmetic prompt rule
• –METEOR plus ROUGE-L emerged as the most reliable reward mix; BLEU alone was not a strong standalone signal for this summarization task
• –The failure mode is familiar: unconstrained quality rewards drift into a coverage-versus-conciseness tradeoff, and the 64-token cap acts like a regularizer
• –The infra is the other noteworthy part: MLX on Apple Silicon plus asynchronous remote rollouts via vLLM is a practical pattern for small teams without a GPU cluster
• –Full bf16 parameters, frozen ref model overhead, and memory-tight training make this a good reference for what is barely feasible on consumer hardware