{"success":true,"data":{"author":"TAMU DIVE Lab (Cong Fu, Shuiwang Ji et al.) with Maho Nakata (RIKEN)","bucketAccessType":"public","bucketName":"divelab/PubChemQCR","bucketPrefix":"","category":"Chemistry & Materials","cloudProvider":"huggingface","createdAt":"2026-11-14T14:15:22.904Z","description":"PubChemQCR, from the TAMU DIVE Lab (Shuiwang Ji and colleagues, with Maho Nakata of RIKEN), contains the geometry relaxation trajectories of about 3.5 million small molecules from the PubChemQC project, optimised sequentially with PM3, Hartree-Fock and DFT, for a total that the authors put at about 300 million molecular conformations with energies and forces. It is released with a benchmark for machine learning interatomic potentials on quantum chemistry relaxations (arXiv:2506.23008). Measured 2026-11-14, the repository root holds LMDB shards Data01.lmdb to Data12.lmdb plus test.lmdb (most around 50 GB each), a Code/ folder with a loader and LMDB structure notes, and uses about 430 GB in total. The loader selects an optimisation stage (pm3, hf, first and second DFT substages, or mixing) and whether to return whole trajectories; read the LMDB files with the python lmdb package and the provided code. The card states that, authorised by Maho Nakata, the processed data are released by TAMU DIVE Lab under CC-BY-4.0.","displayTitle":"PubChemQCR: Relaxation Trajectories of 3.5 Million Molecules for MLIP Benchmarking","id":"2ae84b1e-a514-5f78-a2dc-c02ba1937db0","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/divelab/PubChemQCR","objectCountEstimate":25,"probeStatus":"ok","referenceUrl":"https://arxiv.org/abs/2506.23008","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"divelab/PubChemQCR\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"divelab/PubChemQCR\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=divelab, license:cc-by-4.0, downloads=4169"],"format":"mixed","generatedAt":"2026-11-14T14:15:28.102Z","layout":{"example":"Code/data.py","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"import os\nimport random\nfrom concurrent.futures import ProcessPoolExecutor\nfrom functools import partial\nfrom pathlib import Path\nfrom typing import Optional\nimport torch.nn.functional as F\nimport lmdb\nimport gzip\nimport pickle\nimport json\n\nfrom itertools import product\n\nimport numpy as np\nfrom tqdm import tqdm\nimport torch\n\nfrom torch_geometric.data import Data, Dataset\nfrom torch_geometric.loader import DataLoader\nimport periodictable\nfrom torch.utils.data.distributed import DistributedSampler\nfrom torch.","encoding":"utf8","key":"Code/data.py","truncatedAt":512}],"sampleObjects":["Code/data.py","Code/database.md","Code/lmdb_access.py","Code/main.py","Code/models/painn.py"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\n---\n# PubChemQCR\n\n<p align=\"center\">\n  <img src=\"images/dataset_image.png\" width=\"450\">\n</p>\n\n### Description\nPubChemQCR dataset contains the relaxation trajectory of ~3.5 million small molecules, which can facilitate the development of machine learning interatomic potential (MLIP) models. The relaxation is performed sequentially using PM3, Hartree-Fock, and DFT methods, resulting in a total of 300 million snapshots, 105 million of which are computed using DFT. The dataset is split into two portions, a subset and a full set. Both sets share the same test set but have unique training and validation sets.\n\n\nMore details about the dataset and benchmark can be found in the preprint:\n\n*Cong Fu, Yuchao Lin, Zachary Krueger, Wendi Yu, Xiaoning Qian, Byung-Jun Yoon, Raymundo Arróyave, Xiaofeng Qian, Toshiyuki Maeda, Maho Nakata, Shuiwang Ji:\nA Benchmark for Quantum Chemistry Relaxations via Machine Learning Interatomic Potentials*, [[Link]](https://arxiv.org/abs/2506.23008)\n\n\n### Citation\n\nIf you use this work, please cite:\n```\n@article{fu2025benchmark,\n  title={A Benchmark for Quantum Chemistry Relaxations via Machine Learning Interatomic Potentials},\n  author={Fu, Cong and Lin, Yuchao and Krueger, Zachary and Yu, Wendi and Qian, Xiaoning and Yoon, Byung-Jun and Arr{\\'o}yave, Raymundo and Qian, Xiaofeng and Maeda, Toshiyuki and Nakata, Maho and Ji, Shuiwang},\n  journal={arXiv preprint arXiv:2506.23008},\n  year={2025}\n}\n```\n\n### Data Loading\n#### Flags\n- 'root' : Path to directory containing LMDB files ([Structures of LMDB files](https://huggingface.co/datasets/divelab/PubChemQCR/blob/main/Code/database.md))\n- 'stage' : Which optimization stage data to load\n    - \"pm3\" : Load PM3 data\n    - \"hf\" : Load HF data\n    - \"1st\" : Load DFT first substage data calculated with Firefly/SMASH\n    - \"1st_smash\" : Load only the DFT first substage data calculated with SMASH\n    - \"2nd\" : Load DFT second substage data calculated with GAMESS\n    - \"mixing\" : Load DFT first & second substage data\n\n    \n- 'total_traj' : If true the entire trajectory of a molecule is loaded\n- 'SubsetOnly' : If true then only the subset is loaded\n\n#### Dataset Loading\n```python\nfrom data import LMDBDataLoader, _STD_ENERGY, _STD_FORCE_SCALE\n\nroot = '/path/to/lmdb/dir'\nbatch_size = 128\nnum_workers = 16\nstage = '1st'\ntotal_traj = True\nSubsetOnly = True\n\nloader = LMDBDataLoader(root=root, batch_size=batch_size, num_workers=num_workers, stage=stage, total_traj=total_traj, SubsetOnly=SubsetOnly)\n\ntrain_set = loader.train_loader()\nval_set = loader.val_loader()\ntest_set = loader.test_loader()\n```\n### Training\n#### Important\n- Full dataset training requires a few model functionality to work. See example models for indepth usage.\n- Some molecules have atoms not connected if a cutoff is too small, these nodes need to be removed with 'torch_geometric.utils.remove_isolated_nodes'\n    - Example usage:\n        ```python \n        edge_index, _, mask = remove_isolated_nodes(edge_index, num_nodes=data.num_nodes)\n        pos = data.pos[mask]\n        z = data.x[mask]\n        batch = data.batch[mask]\n        ```\n        - num_nodes flag is needed as without it the function may infer a smaller number of atoms which will cause an error\n- The original batch size needs to be saved for all scatter operations. In rare instances, the whole molecule is removed and passing the original batch size into the scatter function will ensure that molecule gets the value 0. Without it you will get errors.\n    - Example usage:\n        ```python\n            batch_size = data.batch.max().item() + 1\n        ```\n    \n        ```python\n            out = scatter(h, batch, dim=0, dim_size=batch_size, reduce='sum').squeeze()\n        ```\n- Some models require gradient norm clipping in order to prevent loss explosion for some samples. I found gradient clipping to 1.0 was sufficient, but potential clipping values were not thoroughly explored\n    - The log for some epochs may show high losses in the force training phase due to single conformer explosion. \n    - These high losses are because a","encoding":"utf8","key":"README.md","name":"README.md","size":6577,"truncated":true}],"sourceObjectKey":"Code/data.py","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"divelab/PubChemQCR\", filename=\"Code/data.py\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"Code/database.md","totalObjects":25},"summary":"PubChemQCR, from the TAMU DIVE Lab (Shuiwang Ji and colleagues, with Maho Nakata of RIKEN), contains the geometry relaxation trajectories of about 3.5 million small molecules from…","title":"PubChemQCR: Relaxation Trajectories of 3.5 Million Molecules for MLIP Benchmarking","updatedAt":"2026-11-14T17:50:09.305Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/divelab/PubChemQCR"}}