ml-fairchem-finetune

What problem does it solve?

Fine-tuning a foundation Fairchem machine learning interatomic potential to accurately match a specific chemical system or target property without manually wiring an end-to-end training pipeline.

Core Features & Use Cases

• Dataset preparation from JSON labels: Converts structure dictionaries into extxyz and builds Fairchem-compatible LMDB datasets (including generating training configuration templates).
• Foundation model fine-tuning: Runs Fairchem training for UMA/ESEN-style foundation potentials using a generated Hydra configuration and optional backbone freezing.
• Benchmarking and training validation: Benchmarks against a foundation model via the referenced ml-mlip-benchmark skill, then parses logs to confirm convergence and training curves.
• Model registration for reuse: Registers the newly fine-tuned checkpoint into a local registry so future research tasks can discover and reuse it.

Quick Start

Run the data preparation script for your labeled JSON dataset, specifying your Fairchem base model and output directory, so the Skill generates the LMDBs and the Fairchem fine-tuning configuration for training.