Language models can generate VASP INCAR files from natural‑language prompts, yet only large proprietary cloud models have so far handled the tightly coupled, physics‑sensitive settings reliably. This creates a mismatch with local high‑throughput materials workflows that require privacy, low cost, and offline deployment. We demonstrate that a small language model (SLM) can close this gap.
The SLM is fine‑tuned on reference VASP calculations and paired with VASPGuard, a deterministic post‑processor that validates syntax, workflow, and material‑dependent constraints. Together they form the INCAR‑SLM system.
On INCARBench, a benchmark dedicated to VASP INCAR generation, INCAR‑SLM built on Qwen3‑4B outperforms every evaluated general‑purpose LLM, surpassing GPT‑5.4 by 15.55 points on the 100‑point INCAR Score. Most of the improvement stems from fine‑tuning, while VASPGuard corrects the remaining errors.
Additional experiments reveal that model size matters less than expected: after fine‑tuning and post‑processing, performance saturates at a few billion parameters, and Qwen3‑4B even outperforms larger models in the same family.
Review