CHANGELOG: v6 entry wording (initial/preliminary v0)

CHANGELOG.md

@@ -15,32 +15,17 @@ Newest release on top.
 
 ## v6 — no UQ — *2026-09-29* (initial r²SCAN fine-tunes, v0)
 
-- **Initial fine-tune (v0) of the 2L-OMAT models on the MatPES-r²SCAN and MP-ALOE-r²SCAN
-  datasets**, trained by **Quentin Bizot**. Two new models, each as fp32 (default) + `-fp64`:
+- **Initial (preliminary, v0) fine-tune of the 2L-OMAT models on the MatPES-r²SCAN and
+  MP-ALOE-r²SCAN datasets**, trained by **Quentin Bizot**. Two new models, each as fp32
+  (default) + `-fp64`:
   - **`GRACE-2L-OMAT-large-ft-mpaloe-r2scan-v0`** — `GRACE-2L-OMAT-large-base` fine-tuned on
-    MP-ALOE (r²SCAN), 4 epochs. Test MAE 53 meV/atom (energy), 0.096 eV/Å (forces). The energy
-    error was still falling steeply when training stopped; a longer fine-tune will follow.
-  - **`GRACE-2L-OMAT-ft-matpes-r2scan-v0`** — the small legacy `GRACE-2L-OMAT` fine-tuned on
-    MatPES (r²SCAN), 297 epochs (stopped at minimum learning rate). Test MAE 21 meV/atom, 0.100 eV/Å.
-- **r²SCAN total energies.** Both fits used `reference_energy: 0`, so they predict VASP r²SCAN
-  total energies. Do not mix their energies with the PBE-based models.
+    MP-ALOE.
+  - **`GRACE-2L-OMAT-ft-matpes-r2scan-v0`** — `GRACE-2L-OMAT` fine-tuned on MatPES.
+- r²SCAN level of theory.
 - **No UQ.** The checkpoint archives hold only `model.yaml` + `checkpoint.*` (no
   `gmm_artifacts.npz`), and the tags carry no `-uq-vN` suffix.
-- **Kokkos:** `kokkos/GRACE-2L-OMAT-large-ft-mpaloe-r2scan-v0-kokkos.npz` only (shared by both
-  precisions). The small legacy 2L has no kokkos export.
-- SavedModels were re-exported from the training checkpoints (`checkpoint.best_test_loss`, optimizer
-  state stripped) with the current code: `compute`, `compute_dense`, `compute_energy`, plus the 2L
-  parallel split for the large model. The small legacy 2L has a single communicated key, so like
-  its base `GRACE-2L-OMAT` it has no parallel split (`parallel: False` in the client).
-- Gates (2026-09-29): re-export vs the trainer's `final_model` agrees to ≤2e-10 (E/atom, F,
-  stress, fp64); fp32 vs fp64 ≤5e-7 eV/atom, ≤1.1e-6 eV/Å; LAMMPS `grace/2l/kk` vs `pair_style
-  grace` (fp64, Fe–Ni–O, 64 atoms) dE 2e-10 eV/atom, dF 4e-10 eV/Å.
-- **Known limitation: short-range holes.** Under strong isotropic compression the energy turns
-  over and drops below the equilibrium value, as it does in the base models: e.g. for the MP-ALOE
-  model at nearest-neighbour distances below ~1.1 Å (C, diamond), ~1.85 Å (W, bcc), ~1.7 Å
-  (Zn, hcp). The MatPES model shows a new one in diamond C below ~1.25 Å that its base does not
-  have. These distances are far outside equilibrium, but high-temperature or high-pressure MD
-  can reach them.
+- **Kokkos:** `kokkos/GRACE-2L-OMAT-large-ft-mpaloe-r2scan-v0-kokkos.npz` (shared by both
+  precisions of the large model).
 - Tags: `model-v6`, `checkpoint-v6`, `kk-v6`. Earlier tags are untouched.
 
 ---