Fine-tuning

Every framework fine-tunes differently: MACE takes command-line flags, NequIP and SevenNet read YAML configs, DeePMD reads input.json, and each expects its own dataset layout. oh-my-mlip converts your data to that framework's format, writes the framework's own config and command, and runs the upstream trainer — nothing is reimplemented.

Ask your LLM

Fine-tune MACE on frames.traj — just a quick check that fine-tuning works.

Fine-tune SevenNet-MF-OMPA on my DFT frames for 100 epochs.

What you prepare

• A dataset that ASE reads (extxyz, .traj, OUTCAR, …) with energies and forces;
• the model to start from; which models can be fine-tuned is listed in Fine-tuning per framework.

What it asks you

Your agent first reads which settings that framework really has — they differ: DeePMD counts steps instead of epochs, MACE and SevenNet switch stress on through its weight, and not every trainer has EMA or early stopping. Then it asks:

• the exact variant, e.g. MACE-MH-1-OMAT;
• where energies and forces are stored, if they are not on a calculator;
• whether to keep the defaults it shows for training length, batch size, learning rate, energy/force/stress weights and stress on or off, or change any of them (a setting upstream gives no value for must be set, e.g. DPA4's learning rate);
• the output directory.

Every value comes from you, from the framework's official fine-tuning example, or from its upstream default, and the run records which one it was.

What you get

• the converted dataset, the framework's config and training script, and ft_run.json recording what was run;
• the fine-tuned checkpoint, checked to reload and compute energy and forces;
• the license of the starting checkpoint — a model fine-tuned from a non-commercial checkpoint inherits its terms.

Run it yourself

python scripts/ft_run.py MACE --show-settings      # the settings this framework has
python scripts/ft_run.py MACE --dataset frames.traj --out ft_mace --epochs 50 --seed 0
python scripts/ft_verify.py ft_mace/<checkpoint> --model MACE --json

Option	Effect
--show-settings	list native settings, the flag that sets each (if any), default, official fine-tuning value and whether they are required
--epochs, --max-steps, --batch-size, --lr	training length, batch size, learning rate
--energy-weight, --force-weight, --stress-weight	loss weights
--include-stress, --no-stress	train on stress or not
--set NAME=VALUE	any other native setting, by its name in --show-settings
--version MACE-MH-1-OMAT	start from a specific checkpoint
--emit-only	write the dataset, config and command without running
--slurm --partition <p>	also write a SLURM script (nothing is submitted). Its header sets only the job name, partition, log files, one task and one GPU; add wall time, memory, CPUs or account if your cluster's defaults do not suit
--allow-partial-seed	NequIP/Allegro only: their trainers fix part of the seed internally

Convert a dataset without training: python scripts/ft_dataset.py --input frames.traj --to mace --out ft_data. Full procedure: recipes/finetune.md.