We released the LeMat-Rho charge density dataset (Entalpic/lemat-rho-v1, 69,117 r2SCAN densities, 15x15x15 grids in the default config), but this repository carries only the VASP workflow scripts that produced it. The only training code ever proposed upstream is PR #8, which fine tunes ChargE3Net on Jean Zay against the legacy 10x10x10 chunks; it predates the AMD port and the multi arm model comparison, and our Jean Zay hours are exhausted.
In practice the working pipeline lives on a fork branch, and that cost us concretely this summer. The CINES 30 day scratch purge on Adastra deleted the upstream DeepDFT and charge3net clones and the training venv; the DeepDFT run (job 5004725) was OOM killed at step 0 validation because the vendored collate builds a quadratic probe probe neighbor list in small periodic cells; and reconstructing the setup in July required forensic work through rescue copies because none of it was upstream.
We want the training arms (ChargE3Net fine tune, DeepDFT from scratch, a SALTED style invariant baseline, and BOA) merged here together with the Adastra submit scripts, the SCF speedup driver, and their tests, so any teammate can train against the HF dataset without archaeology on personal forks.
Environment: dataset Entalpic/lemat-rho-v1 (default config, 2026-06-09 revision); Adastra MI250, torch 2.13.0+rocm7.2, python 3.11; reference jobs 5003890 (ChargE3Net), 5004725 (DeepDFT OOM), 5240525/5240531 (current runs).
We released the LeMat-Rho charge density dataset (Entalpic/lemat-rho-v1, 69,117 r2SCAN densities, 15x15x15 grids in the default config), but this repository carries only the VASP workflow scripts that produced it. The only training code ever proposed upstream is PR #8, which fine tunes ChargE3Net on Jean Zay against the legacy 10x10x10 chunks; it predates the AMD port and the multi arm model comparison, and our Jean Zay hours are exhausted.
In practice the working pipeline lives on a fork branch, and that cost us concretely this summer. The CINES 30 day scratch purge on Adastra deleted the upstream DeepDFT and charge3net clones and the training venv; the DeepDFT run (job 5004725) was OOM killed at step 0 validation because the vendored collate builds a quadratic probe probe neighbor list in small periodic cells; and reconstructing the setup in July required forensic work through rescue copies because none of it was upstream.
We want the training arms (ChargE3Net fine tune, DeepDFT from scratch, a SALTED style invariant baseline, and BOA) merged here together with the Adastra submit scripts, the SCF speedup driver, and their tests, so any teammate can train against the HF dataset without archaeology on personal forks.
Environment: dataset Entalpic/lemat-rho-v1 (default config, 2026-06-09 revision); Adastra MI250, torch 2.13.0+rocm7.2, python 3.11; reference jobs 5003890 (ChargE3Net), 5004725 (DeepDFT OOM), 5240525/5240531 (current runs).