docs: clarify LoRA scope and update DPA-3 citation

Author: njzjz

CITATIONS.bib

@@ -224,18 +224,17 @@ @article{Zhang_npjComputMater_2024_v10_p293
   doi          = {10.1038/s41524-024-01493-2},
 }
 
-@misc{Zhang_2025_DPA3,
+@article{Zhang_npjComputMater_2026_DPA3,
   annote       = {DPA-3 descriptor (dpa3)},
   author       = {
     Duo Zhang and Anyang Peng and Chun Cai and Wentao Li and Yuanchang Zhou and
     Jinzhe Zeng and Mingyu Guo and Chengqian Zhang and Bowen Li and Hong Jiang
     and Tong Zhu and Weile Jia and Linfeng Zhang and Han Wang
   },
   title        = {{A Graph Neural Network for the Era of Large Atomistic Models}},
-  publisher    = {arXiv},
-  year         = 2025,
-  doi          = {10.48550/arXiv.2506.01686},
-  url          = {https://arxiv.org/abs/2506.01686},
+  journal      = {npj Comput. Mater.},
+  year         = 2026,
+  doi          = {10.1038/s41524-026-02146-2},
 }
 
 @misc{Li_2026_DPA4,

README.md

@@ -20,9 +20,9 @@ and deploy it at simulation scale.**
 
 > [!IMPORTANT]
 > **A pretrained model can be your starting point, not just your end result.**
-> Download a built-in DPA checkpoint, fine-tune the full model or a LoRA
-> adapter on system-specific data, then test, export, and deploy it through the
-> same DeePMD-kit workflow.
+> Download a built-in DPA checkpoint, fine-tune the full model, or use a
+> [DPA-4 LoRA adapter][dpa4-lora] with PyTorch single-task training, then test,
+> export, and deploy it through the same DeePMD-kit workflow.
 
 DeePMD-kit turns quantum-mechanical reference data into fast, scalable
 interatomic potentials. Use it across molecular and materials science—from
@@ -38,16 +38,16 @@ dynamics.
 
 ## ⚡ Why DeePMD-kit
 
-|     | Advantage                           | What it unlocks                                                                                                                                                                  |
-| --- | ----------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| 🧬  | **Pretrained-first workflows**      | Download [pretrained DPA models][pretrained], fine-tune full models or LoRA adapters, or adapt learned representations to downstream properties with [DPA-ADAPT].                |
-| 🏗️  | **Training from scratch**           | Design a model for a new system or physical target, then train it with single-task, multi-task, and distributed workflows across supported backends.                             |
-| 🧠  | **Modern model portfolio**          | Start with efficient DeepPot-SE descriptors or move to [DPA][model-guide] for large atomistic models.                                                                            |
-| 🧲  | **More than energy and force**      | Model virials, Hessians, spin and magnetic forces, dipoles, polarizabilities, electronic density of states, atomic populations, and arbitrary intensive or extensive properties. |
-| 🔄  | **Backend flexibility**             | Train or run supported models with [TensorFlow, PyTorch, JAX, or Paddle][backends], with backend-aware model formats and conversion paths for compatible architectures.          |
-| 🚀  | **Performance from training to MD** | Use CPUs, CUDA GPUs, ROCm source builds, distributed training, model compression, compiled DPA-4 paths, AOTInductor `.pt2` export, and MPI-enabled simulation.                   |
-| 🔌  | **Deploy where science happens**    | Use the CLI, Python, C, C++, or Node.js, then connect models to LAMMPS, i-PI, ASE, GROMACS, JAX MD, nvalchemi, OpenMM, Amber, CP2K, ABACUS, and more.                            |
-| 🧩  | **Open and extensible**             | Compose hybrid potentials, add analytical ZBL or long-range corrections, create custom models and operators, or connect external GNNs such as MACE and NequIP through plugins.   |
+|     | Advantage                           | What it unlocks                                                                                                                                                                                                           |
+| --- | ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🧬  | **Pretrained-first workflows**      | Download [pretrained DPA models][pretrained], fine-tune full models, use [DPA-4 LoRA adapters][dpa4-lora] with PyTorch single-task training, or adapt learned representations to downstream properties with [DPA-ADAPT]. |
+| 🏗️  | **Training from scratch**           | Design a model for a new system or physical target, then train it with single-task, multi-task, and distributed workflows across supported backends.                                                                      |
+| 🧠  | **Modern model portfolio**          | Start with efficient DeepPot-SE descriptors or move to [DPA][model-guide] for large atomistic models.                                                                                                                     |
+| 🧲  | **More than energy and force**      | Model virials, Hessians, spin and magnetic forces, dipoles, polarizabilities, electronic density of states, atomic populations, and arbitrary intensive or extensive properties.                                          |
+| 🔄  | **Backend flexibility**             | Train or run supported models with [TensorFlow, PyTorch, JAX, or Paddle][backends], with backend-aware model formats and conversion paths for compatible architectures.                                                   |
+| 🚀  | **Performance from training to MD** | Use CPUs, CUDA GPUs, ROCm source builds, distributed training, model compression, compiled DPA-4 paths, AOTInductor `.pt2` export, and MPI-enabled simulation.                                                            |
+| 🔌  | **Deploy where science happens**    | Use the CLI, Python, C, C++, or Node.js, then connect models to LAMMPS, i-PI, ASE, GROMACS, JAX MD, nvalchemi, OpenMM, Amber, CP2K, ABACUS, and more.                                                                     |
+| 🧩  | **Open and extensible**             | Compose hybrid potentials, add analytical ZBL or long-range corrections, create custom models and operators, or connect external GNNs such as MACE and NequIP through plugins.                                            |
 
 > [!TIP]
 > On supported descriptors and workloads, [model compression][compression] can
@@ -76,9 +76,10 @@ flowchart LR
    adaptation, or configure a model to train from scratch.
 1. **Prepare target data** in DeePMD's NumPy format or convert structures and
    trajectories with [dpdata][data].
-1. **Fine-tune or train:** adapt the full pretrained model or LoRA adapters,
-   or optimize a new model with single-task, multi-task, and distributed
-   training workflows.
+1. **Fine-tune or train:** adapt the full pretrained model, use
+   [DPA-4 LoRA adapters][dpa4-lora] with PyTorch single-task training, or
+   optimize a new model with single-task, multi-task, and distributed training
+   workflows.
 1. **Validate and export** with [`dp test`][testing], [`dp freeze`][freeze],
    backend conversion, embedding extraction, and supported compression paths.
 1. **Run simulation** through Python or native APIs, or load the model into a
@@ -119,7 +120,8 @@ mode with `dp --pt` and select the branch that matches your system with
 configuration, so `input.json` does not need to reproduce the DPA-3.2
 architecture.
 
-The [fine-tuning guide][finetune] covers full-model and LoRA adaptation.
+The [fine-tuning guide][finetune] covers full-model adaptation. [DPA-4 LoRA
+fine-tuning][dpa4-lora] is available for PyTorch single-task training.
 [DPA-ADAPT] reuses pretrained DPA representations for downstream
 property-prediction tasks.
 
@@ -256,6 +258,7 @@ DeePMD-kit is licensed under the
 [data]: https://docs.deepmodeling.com/projects/deepmd/en/latest/data/dpdata.html
 [documentation]: https://docs.deepmodeling.com/projects/deepmd/en/latest/
 [dpa-adapt]: https://docs.deepmodeling.com/projects/deepmd/en/latest/dpa_adapt/overview.html
+[dpa4-lora]: https://docs.deepmodeling.com/projects/deepmd/en/latest/model/dpa4.html#lora-fine-tuning
 [embeddings]: https://docs.deepmodeling.com/projects/deepmd/en/latest/inference/embedding.html
 [finetune]: https://docs.deepmodeling.com/projects/deepmd/en/latest/train/finetuning.html
 [freeze]: https://docs.deepmodeling.com/projects/deepmd/en/latest/freeze/freeze.html

doc/index.rst

@@ -10,9 +10,9 @@ DeePMD-kit
 .. important::
 
    **A pretrained model can be your starting point, not just your end result.**
-   Download a built-in DPA checkpoint, fine-tune the full model or a LoRA
-   adapter on system-specific data, then test, export, and deploy it through
-   the same DeePMD-kit workflow.
+   Download a built-in DPA checkpoint, fine-tune the full model, or use
+   `DPA-4 LoRA`_ with PyTorch single-task training, then test, export, and deploy
+   it through the same DeePMD-kit workflow.
 
 DeePMD-kit turns quantum-mechanical reference data into fast, scalable
 interatomic potentials. Use it across molecular and materials science—from
@@ -46,7 +46,8 @@ Choose your path
       :link-type: doc
       :shadow: md
 
-      Adapt a full pretrained model or LoRA adapters to your target system.
+      Adapt a full pretrained model, or use DPA-4 LoRA with PyTorch single-task
+      training.
 
    .. grid-item-card:: 🏗️ Train from scratch
       :link: train/index
@@ -81,8 +82,9 @@ Why DeePMD-kit
    .. grid-item-card:: Pretrained-first adaptation
       :shadow: sm
 
-      Start from built-in DPA checkpoints, fine-tune full models or LoRA
-      adapters, and reuse learned representations with DPA-ADAPT.
+      Start from built-in DPA checkpoints, fine-tune full models, use
+      `DPA-4 LoRA`_ with PyTorch single-task training, and reuse learned
+      representations with DPA-ADAPT.
 
    .. grid-item-card:: Training from scratch
       :link: train/index
@@ -156,7 +158,7 @@ Two starting points, one path to dynamics
       :link: train/finetuning
       :link-type: doc
 
-      Adapt the full model or LoRA adapters to your system.
+      Adapt the full model, or use DPA-4 LoRA with PyTorch single-task training.
 
    .. grid-item-card:: 3 · Train
       :link: train/index
@@ -177,7 +179,8 @@ Two starting points, one path to dynamics
       Run inference directly or deploy into molecular dynamics.
 
 Fine-tuning and from-scratch training converge on the same validation, export,
-and deployment toolchain.
+and deployment toolchain. `DPA-4 LoRA`_ is currently limited to PyTorch
+single-task fine-tuning.
 
 Choose a model family
 =====================
@@ -200,8 +203,9 @@ Choose a model family
      - Message passing over line-graph representations and broad chemical
        coverage.
    * - :doc:`DPA-4 <model/dpa4>`
-     - SO(3)-equivariant learning, LoRA fine-tuning, optional ZBL bridging,
-       spin support, and compiled ``.pt2`` deployment.
+     - SO(3)-equivariant learning, `DPA-4 LoRA`_ for PyTorch single-task
+       fine-tuning, optional ZBL bridging, spin support, and compiled ``.pt2``
+       deployment.
 
 .. figure:: _static/dpa4-performance.webp
    :alt: DPA4 energy and force accuracy versus saturated throughput
@@ -256,8 +260,8 @@ New and noteworthy
       :link-type: doc
       :shadow: sm
 
-      Equivariant message passing, LoRA, ZBL, spin, compiled inference, and
-      LAMMPS deployment.
+      Equivariant message passing, PyTorch single-task LoRA, ZBL, spin,
+      compiled inference, and LAMMPS deployment.
 
    .. grid-item-card:: DPA-ADAPT
       :link: dpa_adapt/index
@@ -425,5 +429,7 @@ Indices
 * :ref:`modindex`
 * :ref:`search`
 
+.. _DPA-4 LoRA: https://docs.deepmodeling.com/projects/deepmd/en/latest/model/dpa4.html#lora-fine-tuning
+
 .. _feedback:
 .. _affiliated packages: