Autonomous Weighted Schema Direct Preference Optimization for Materials Discovery (Dataset and Code)
Abstract
NOTE: Metadata and Authors will be updated upon de-anonymization by conference. Title: Autonomous Weighted Schema Direct Preference Optimization for Materials Discovery (dataset)Authors: REDACTED Files for post-training of Pharmacological discovery models using supervised fine-tuning (SFT) and reinforcement learning (RL). /configurations: Configuration files for training. Absolute paths have been anonymized and should be adapted for reproduction efforts. Example: preloaded_model_path: #see https://zenodo.org/records/19374557 tokenizer_path: #see https://zenodo.org/records/19374557 dset_path: /Databases/Pharma/safe_subset_qed_greq_pt7.parquet properties_name: safe_properties.json method: base dset_type: pharma save_dir_base: /saved_models/PropXGPT_FT/large_group_pharma/bpe1000/base version: 1 model_index: 2 nepochs: 5 gen_nsamples: 8192 valits: 200 lr: 0.0001 suffix: bpe_vocab_1000 nnodes: 1 /dataset/pharma_sft: Fine-tuning dataset for pretraining. This is a downsampling of https://huggingface.co/datasets/datamol-io/safe-gpt (Apache 2.0) with additional tokenization processing and determined values for molecules. Example data: level_0 1 input CC6.[C@@H]16CN5C[C@]2(CCOC2)O1.C54=O.c14cnn2cc... mc_labels [315.15829153199996, 0.5294117647058824, 6.0, ... smiles CC[C@@H]1CN(C(=O)c2cnn3ccccc23)C[C@]2(CCOC2)O1 selfies [C][C][C@@H1][C][N][Branch2][Ring1][C][C][=Bra... group_selfies_bpe_vocab_1000 [:0CC][C][:0Cchi3][Ring1][:0NbraCketCbraeqOket... index 0 qed 0.846995 sa_score 4.09037 n_heavy_atoms 23.0 n_rings 4.0 /modelsModel weights and associated files. Model weights (*ckpt), training and validation losses (*npy), and benchmarking data (*csv, *parquet). See MODELS_LICENSE.md for models licensing information (MIT). Over the course of DPO training, samples are paired into preferences (e.g., Pref.parquet). Final DPO loops include final model configuration.For storage contraints and due to the presence of replicates, only a representative model (e.g., "i0") of the models considered in the main text and the initial SFT model are included here. Example data:Samples: raw_group_selfies [START][:0CNC][C][:2coneccnccone][Branch1][:0C... group_selfies [:0CNC][C][:2coneccnccone][Branch1][:0Cl][pop]... canon_smiles Nc1cc([O-])nc(Cl)c1 Benchmarks: selfies_validity 1.000000 uniqueness 1.000000 synthetic_accessibility 3.651891 synthetic_inaccessibility 6.348109 oxygen_balance_hundred -84.615385 qed 0.534761 n_nn_bonds 0.000000 n_no_bonds 0.000000 n_oo_bonds 0.000000 n_so_bonds 0.000000 n_halogens 1.000000 n_heavy_atoms 9.000000 n_radicals 0.000000 n_rings 1.000000 n_aromatic_rings 1.000000 n_fused_rings 0.000000 bcut2D_mw_low 10.409414 bcut2D_mw_high 35.495696 bcut2D_mr_low -1.951055 bcut2D_mr_high 6.292152 bcut2D_chg_low -2.029981 bcut2D_chg_high 1.875222 bcut2D_logP_low -1.951055 bcut2D_logP_high 2.129304 fn:qed-.001*(n_heavy_atoms-20)**2 0.413761 fn:synthetic_accessibility-.01*(n_heavy_atoms-20)**2 2.441891 /plots Plotting scripts. Running ```pythonpython plots.py``` with the activated environment from the supplementary codebase will generate plots. /schemas File(s) for initializing training objects. Load with supplementary codebase.
