About FinetuneLLM — who builds it and how the estimates work

Peter Peeters

I build and run software, and I got tired of the same failure mode: renting a GPU, waiting for the dataset to tokenise, and hitting an out-of-memory error forty steps in. The information needed to avoid that is not secret. It is arithmetic, and almost nobody publishes it in a form you can apply to your own model.

FinetuneLLM is that arithmetic, made usable. Every number comes from a model I publish in full and argue with in public.

github.com/peterpeeterspeter

How the estimates are built

Six terms, each derived from the model architecture rather than from a benchmark table:

weights      = params x bytes_per_param     (2 bf16 · 0.55 NF4)
gradients    = trainable x 2
optimiser    = trainable x 10               (AdamW fp32 m, v, master)
activations  = batch x seq x hidden x layers x 2 x ckpt_factor
logits       = batch x seq x vocab x 2 x ce_factor
overhead     = 0.8 GB

Throughput uses a 6·N flops-per-token baseline with a measured efficiency factor per method. That is the weakest part of the cost estimate, and it is why the range you should trust is ±20-40% rather than the two decimal places the calculator prints.

Where the numbers come from

Model architectures (layers, hidden size, vocabulary) come from the published configs on Hugging Face. GPU prices are marketplace medians for October 2026 — they move weekly, and the calculator lets you overwrite them with what you actually pay.

Search volumes quoted in the guides are Keyword Planner figures for the US, English, pulled on the date shown.

Cross-checks

The model is only worth something if it matches reality. I validate it against real runs and publish the delta here, including the ones where it is wrong. This table is intentionally empty at launch rather than filled with estimates dressed up as measurements — the first entries go up once the runs complete.

Run	Estimated	Measured peak	Delta
Pending — first validation runs in progress.

Corrections

If an estimate is wrong for your setup, that is useful. Open an issue on the GitHub profile above with the model, the settings and the peak memory you saw, and it gets fixed and credited.