Autonomous LLM post-training with Tunix on TPUs
The article introduces autofinetune, a system applying autonomous AI research loops to LLM post-training (SFT and RL via GRPO) using Google’s stack—Tunix, Gemma, and Cloud TPUs with Antigravity CLI and Gemini Flash 3.7. Inspired by autoresearch, it automates the manual finetuning cycle: humans define boundaries (program.md), agents iterate on a training script (run.py), measure metrics, and commit improvements to Git.
Case Study 1: SFT on FunctionGemma-270m-it ran 20 experiments in hours on TPU v5e-1, auto-tuning LoRA ranks, learning rates, and optimizers to boost function-call accuracy.
Case Study 2: GRPO on Gemma 3 1B (GSM8K) ran 40 experiments over 2–3 days on TPU v6e-1, improving reward ~10%.
