Koulb commited on
Commit
2b5ba77
·
verified ·
1 Parent(s): db31a41

Add diamond HPRO DeepH release README

Browse files
releases/diamond_hpro_deeph_v2/README.md ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Diamond HPRO/DeepH-E3 Data Release v2
2
+
3
+ This release adds the cluster-generated diamond HPRO/DeepH-compatible labels used for ZPR, EPC, and Raman experiments.
4
+ The old top-level Hugging Face layout is left untouched; this release lives under `releases/diamond_hpro_deeph_v2/`.
5
+
6
+ ## Datasets
7
+
8
+ - `data/zg_endpoints`: 12 frames. Equilibrium and zero-point-renormalized ZG endpoint structures for sc3..sc8, used as the main ZPR reference and validation target.
9
+ - `data/zg_train_a0`: 61 frames. Original 61-frame 3x3x3 primitive-cell diamond training set that gave the best pre-A6 baseline for ZPR.
10
+ - `data/zg_train_a1_clean`: 60 frames. Sixty additional random displacement frames, designed to test whether simply adding more similar frames improves ZPR.
11
+ - `data/zg_train_a2_zgline`: 20 frames. Twenty frames sampled along and near the actual sc3 ZG displacement direction.
12
+ - `data/zg_train_a7_physical_fd`: 64 frames. Sixty-four physical finite-difference frames around ZPR-relevant endpoint and phonon directions.
13
+ - `data/raman_v3_308`: 308 frames. Large 308-frame HPRO dataset for diamond EPC/Raman and Hamiltonian training, spanning small FD, random displacement amplitudes, strain, and sc4 transfer frames.
14
+ - `data/ki_delta_optional`: 60 frames. Diamond KI/Koopmans Hamiltonian-label experiment data, kept separate because it is not part of the primary EPC/ZG/Raman HPRO label family.
15
+
16
+ ## Size
17
+
18
+ - HDF5 files in this release: 1172
19
+ - HDF5 payload size: 42.225 GiB
20
+
21
+ ## Files Included
22
+
23
+ Each frame includes `hamiltonians.h5` and `overlaps.h5` when available, plus lightweight input/metadata files.
24
+ QE scratch directories, wavefunction save trees, DeepH graph pickles, and bulk training checkpoints are intentionally excluded.
25
+
26
+ ## Local Compatibility
27
+
28
+ Use `compatibility/materialize_epc_ml.py` to create a symlink-based local layout after downloading this release.
29
+
30
+ ## Provenance
31
+
32
+ Cluster source paths are recorded in every dataset README and in each frame `meta.release.json`.
33
+ Release staged: 2026-05-18T16:17:13.259371+00:00