LFM2.5-350M wins Reddit summarization evals

This is a solid small-model RL result: the win is not just that the model got better, but that reward shaping clearly mattered more than brute length pressure. It’s still a narrow benchmark, but the infra story is the more interesting part because it shows real fine-tuning experiments running on commodity Apple silicon.

• –The best configuration outperformed length-only training on composite score, faithfulness, coverage, conciseness, and pass rate.
• –METEOR + ROUGE-L was the strongest reward mix, which fits because it captures paraphrase and overlap better than BLEU alone.
• –The 3x Mac mini, MLX, vLLM-metal, SyncPS setup shows heterogeneous local clusters can support meaningful RL fine-tuning work.
• –The evaluation is still limited by a 200-example test set and judge-based scoring, so this reads as promising evidence rather than a definitive benchmark victory.