Training Tiny LLMs for 64‑Token Summaries on a Mac Mini Cluster

The author updated a GRPO training pipeline aimed at producing 64‑token summaries of length‑constrained Reddit posts, using LFM 2.5‑350 M and Qwen 2.5‑0.5 B‑Instruct models. A length penalty was added to force exactly 64 tokens, but this caused BLEU and ROUGE‑L scores to drop when training from scratch, prompting the use of a fine‑tuned checkpoint that already outputs 64 tokens. Evaluation employed DeepEval with a GPT‑5 judge scoring faithfulness, coverage, conciseness, and clarity, and the results were plotted for each model. Training ran on a 3‑node Mac Mini cluster with MLX, where one node drove GRPO and two nodes ran vLLM‑metal rollouts; the setup used a synchronous parameter‑server architecture (SyncPS) with the master node handling training. The distributed training leveraged the smolcluster framework, and the vLLM‑metal framework was used for efficient inference on the worker nodes. The pipeline demonstrated that length‑constrained summarization is feasible on modest hardware, though BLEU/ROUGE‑L remain low without a pre‑trained length‑aware checkpoint. The author plans to extend the work to DPO comparison and further fine‑tuning experiments. The post includes screenshots of training logs and evaluation metrics, illustrating the current performance of the models.