LLaMo: Large Language Model-based Molecular Graph Assistant

Authors: Jinyoung Park, Minseong Bae, Dohwan Ko, Hyunwoo J. Kim

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments demonstrate that LLa Mo shows the best performance on diverse tasks, such as molecular description generation, property prediction, and IUPAC name prediction.
Researcher Affiliation Academia Jinyoung Park Minseong Bae Dohwan Ko Hyunwoo J. Kim Department of Computer Science and Engineering, Korea University {lpmn678, bms2002, ikodoh, hyunwoojkim}@korea.ac.kr
Pseudocode No The paper describes methods in prose and figures but does not include pseudocode or algorithm blocks.
Open Source Code Yes The code of LLa Mo is available at https://github.com/mlvlab/LLa Mo.
Open Datasets Yes For molecule description generation , and property prediction , we use the datasets derived from Pub Chem and QM9 of Molecule Net [64] as in Mol-Instructions [48]. For IUPAC name prediction, a dataset derived from [3] is used. To train the generalist variant of LLa Mo, we use a training split of molecular description generation dataset of Mol-Instructions in stage 1. In stage 2, the model is instruction-tuned with a training split of description generation and property prediction instruction dataset of Mol-Instructions, IUPAC name prediction from [3], and our GPT-generated instruction-following data. [...] Pub Chem324k is constructed by collecting 324k molecules and their associated text information from the Pub Chem database. Ch EBI-20 is the most commonly utilized benchmark in this task, consisting of selected 33,010 pairs of molecules and descriptions from Ch EBI [72].
Dataset Splits Yes To train the generalist variant of LLa Mo, we use a training split of molecular description generation dataset of Mol-Instruction [48] in stage 1. In stage 2, the model is instruction-tuned with a training split of description generation and property prediction instruction dataset of Mol-Instructions, IUPAC name prediction from [3], and our GPT-generated instruction-following data. For the evaluation of molecular description generation and property question answering tasks, we use the test split of Mol-Instructions molecular description generation and property prediction datasets, which are sampled from Pub Chem [44] and QM9 dataset of Molecule Net [64], respectively.
Hardware Specification Yes Our experiments are run on 4 A6000 GPUs or 4 V100 GPUs and 2 A6000 GPUs for LLa MA2 and Galactica, respectively.
Software Dependencies No The paper mentions software like PyTorch, PyTorch Geometric, Huggingface transformers, PEFT, and Open Delta, but does not specify their version numbers.
Experiment Setup Yes In stage 1, the Adam W [63] optimizer is adapted with an initial learning rate of 1e-4 (minimum learning rate is 1e-5 and warmup learning rate is 1e-6). The warmup step is 1,000 and the cosine scheduler is applied. In stage 2, the initial learning rate is set to 5e-5 (minimum learning rate is 5e-6 and warmup learning rate is 5e-7). [...] We use Lo RA to train the large language model in stage 2.