Skip to content

LlaSMol

OSU Natural Language Processing Group

LlaSMol is a family of chemistry large language models fine-tuned on the SMolInstruct instruction dataset (four released models) covering tasks such as name conversion, property prediction and reaction prediction, with the code under the MIT license.

Catalog updated ·

Overview

LlaSMol is the chemistry large language model family from the paper 'LlaSMol: Advancing Large Language Models for Chemistry with a Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset' (arXiv 2402.09391), developed by the Ohio State University NLP group. This repository is the paper's official code release.

The README documents the project page, the SMolInstruct dataset on Hugging Face (osunlp/SMolInstruct), and four released models (LlaSMol-Galactica-6.7B, LlaSMol-Llama2-7B, LlaSMol-CodeLlama-7B, LlaSMol-Mistral-7B), trained on 4 task types and 14 tasks that include name conversion, property prediction and reaction-related tasks, with worked input/output examples in the README.

The repository code is MIT licensed. The four Hugging Face releases contain LoRA adapter weights, not standalone full base-model weights; all four model cards declare CC BY 4.0. SMolInstruct is also explicitly licensed under CC BY 4.0. Loading an adapter requires its corresponding Galactica, Llama2, CodeLlama or Mistral base model, whose separate terms still apply. The model cards describe research use and warn of incorrect outputs. Check generated structures, reactions and property answers against authoritative chemistry sources. Model card: https://huggingface.co/osunlp/LlaSMol-Mistral-7B ; dataset: https://huggingface.co/datasets/osunlp/SMolInstruct

Key Features

  • Instruction-tuned chemistry LLMs released in four variants based on Galactica, Llama2, CodeLlama and Mistral backbones.
  • Trained on SMolInstruct, a large-scale chemistry instruction tuning dataset published on Hugging Face.
  • Covers 4 task types and 14 tasks documented in the README, including name conversion and property prediction.

Use Cases

  • Experiment with instruction-following chemistry assistants for property and reaction questions.
  • Reproduce or extend the paper's fine-tuning pipeline with the released dataset and code.

How to Use

Follow the repository generation example to load a named LlaSMol LoRA adapter together with its corresponding base model; use the documented input tags for SMILES and IUPAC names. Adapter cards declare CC BY 4.0, code is MIT, and SMolInstruct is CC BY 4.0; obtain the base model under its own terms. The releases are research models and outputs require chemical verification. See: https://github.com/OSU-NLP-Group/LLM4Chem

Sources and verification

This entry describes capabilities documented by the linked project sources. Catalog edits are not evidence of a successful run, compatibility test or scientific performance evaluation.

Repository source reviewed for this catalog version:

Read the editorial and verification policy

Related resources

ASKCOS

Platforms

ASKCOS is a scientific synthesis-planning platform with retrosynthesis, forward prediction and reaction-analysis modules, offered through a hosted interface and modular deployment documentation.

Python

Reaction Prediction · Retrosynthesis

Graph2SMILES is a graph-to-sequence model for single-step retrosynthesis and chemical reaction outcome prediction.

Open sourcePython

Reaction Prediction · Retrosynthesis

Molfeat

Open Source

Molfeat is a Python toolkit for small-molecule featurization, combining molecular fingerprints, descriptors and pretrained embeddings with parallel transformation, caching and extensible plugins.

Open sourcePython

Molecular Property Prediction · Cheminformatics

Mordred

Open Source

Mordred is a Python molecular descriptor calculator with command-line and RDKit-based library workflows, producing descriptor tables for cheminformatics analysis and downstream model evaluation.

Open sourcePython

Molecular Property Prediction · Cheminformatics

RxnMapper uses attention from an unsupervised ALBERT model to assign atom mappings to valid reaction SMILES, returning mapped reactions and confidence scores through a Python interface.

Open sourcePython

Cheminformatics · Reaction Prediction

RXN Reaction Preprocessing is a pipeline for preprocessing chemical-reaction datasets, covering standardization, filtering, augmentation, splitting and tokenization.

Open sourcePython

Cheminformatics · Reaction Prediction

Related guides

Models

Choosing between Chemprop and DeepChem

Compare Chemprop’s molecular property prediction focus with DeepChem’s broader scientific machine-learning scope, then plan a fair evaluation using shared data, explicit inputs and outputs, and reproducible decision criteria.