README.md

metadata

language:
  - en
tags:
  - chemistry
  - molecules
  - drug-discovery
  - molecular-generation
  - multimodal
base_model: Qwen/Qwen2.5-3B
library_name: transformers
datasets:
  - language-plus-molecules/mCLM_Pretrain_1k
license: cc-by-nc-nd-4.0

mCLM_1k-3b

mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules

Relevant Links

:globe_with_meridians: Website | :octocat: Code | :hugs: Data and Model | :desktop_computer: Demo | :page_with_curl: Paper

Architecture

• Base Model: Qwen2.5-3B
• Molecular Encoder: GNN with 5 message passing layers
• Molecular Tokenizer: Custom block-based tokenizer for SMILES representations

Usage

Please follow installation instructions from the Github .

from mCLM.model.models import mCLM
from mCLM.tokenizer.utils import convert_instruction_to_input, message_ids_to_string, get_processor
import torch

# ===========================
# Settings
# ===========================
DTYPE = torch.bfloat16
DEVICE = torch.device("cpu")

if __name__ == "__main__":

        
    model = mCLM.from_pretrained("language-plus-molecules/mCLM_1k-3b")

    tokenizer = model.tokenizer
    molecule_tokenizer = model.molecule_tokenizer

    bad_words_ids = None

    model.to(DEVICE).to(DTYPE) #This is important for the HF model
        
    while True:
        user_input = input("Enter an instruction (type 'quit' to exit): ")
        if user_input == 'quit': break
        user_input = user_input.strip()

        message_tokens = convert_instruction_to_input(user_input, model, molecule_tokenizer, tokenizer)
            
        ################## Generate results ###################################

        beam_size = 5            

        input_ids = message_tokens.to(DEVICE)

        processor = get_processor(molecule_tokenizer, tokenizer) #we do this every time in case vocab was expanded

        generated = model.generate(
            input_ids=input_ids,
            attention_mask=torch.ones_like(input_ids), #This is to turn off the attention mask warning
            pad_token_id=tokenizer.eos_token_id, #This is to turn off the pad token warning
            max_new_tokens=32,
            num_beams=beam_size,
            num_return_sequences=beam_size,
            logits_processor=processor,
            do_sample=False,
            bad_words_ids=bad_words_ids,
            diversity_penalty=1.0,
            num_beam_groups=beam_size,
        )

        for i in [0]: #range(beam_size):
            message_ids = generated[i, message_tokens.shape[1]:]
            
            mol_msg, smiles_msg, mol_list, smiles_list = message_ids_to_string(message_ids, molecule_tokenizer, tokenizer)

            if smiles_msg != None:
                print(mol_msg)
                if len(smiles_list) > 0:
                    print("SMILES list:", smiles_list)
            else:
                print(mol_msg)

            print()

Training Data

The model was trained on:

• Molecular instruction-following data from activity cliffs
• General text instruction data, SMolInstruct, Mol-Instructions Biomedical

Citation

If you use this model, please cite:

@misc{edwards2025mclmmodularchemicallanguage,
      title={mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules}, 
      author={Carl Edwards and Chi Han and Gawon Lee and Thao Nguyen and Sara Szymkuć and Chetan Kumar Prasad and Bowen Jin and Jiawei Han and Ying Diao and Ge Liu and Hao Peng and Bartosz A. Grzybowski and Martin D. Burke and Heng Ji},
      year={2025},
      eprint={2505.12565},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2505.12565}, 
}

Model Card Contact

For questions or issues, please open an issue in the repository.