Autonomous LLM post-training with Tunix on TPUs
~2 min read
aiaillmfine-tuningagentsautoresearchtpu
A disciplined agent-eval workflow: the red-pen day, a golden dataset with framework-agnostic OpenTelemetry traces, deterministic assertions plus boolean judges, and 50–60 small tuning iterations instead of full rewrites.
LLM post-training is a repetitive, mostly mechanical discipline: run a training job, watch loss curves, check benchmark evaluations, manually revert failures or record successes in a spreadsheet. The "autofinetune" project asks the obvious question — if agents can do research, why can't they do this loop too?
As demonstrated in autoresearch, the whole process can be automated with AI agents in three parts:
program.md) — the human defines the loop: boundary conditions, evaluation criteria, and constraints.run.py) — a single, clean, self-contained finetuning script.program.md: it modifies run.py, runs the training job, measures the target metric, retains winning commits or reverts regressions, and logs results in results.tsv.The human's job shrinks to defining what good looks like; the agent's job is the loop — propose, run, measure, keep-or-revert, log.
The first experiment in autofinetune took the same Supervised Fine-Tuning (SFT) setup from an earlier post and extended it into the autoresearch loop: optimize google/functiongemma-270m-it on the google/mobile-actions dataset. Rather than one hand-tuned run, the agent explores the training configuration space, keeps what improves the metric, and leaves a reproducible results.tsv trail behind.
This is a genuinely different workflow from "one big fine-tune": post-training becomes a search problem the agent runs overnight. Every candidate config is measured against the same metric, winners and losers are both recorded, and the human reviews the arena rather than babysitting each run. It won't replace research questions — but it does replace the spreadsheet.
If you fine-tune models on TPUs, autofinetune is the pattern to steal: define your boundary conditions once, hand the iteration loop to an agent, and let a reproducible results log accumulate instead of tribal knowledge. The arena design — program.md + run.py + agent loop — is a clean template for any repetitive experimental workflow.