smolgen-pubchem-46M-base

A 46M-parameter causal language model for de novo molecule generation trained on SMILES strings from PubChem.

Training Data

The model was pretrained on ~40 million molecules sourced from PubChem and filtered by:

• Heavy atom count: only drug-like size molecules retained
• Structure alerts: compounds flagged by common medicinal chemistry filters removed
• Salt removal: only the largest fragment of each compound kept

Model Architecture

Decoder-only Transformer (LlamaForCausalLM) with grouped-query attention (GQA):

Parameter	Value
Hidden size	576
Intermediate size	1536
Layers	13
Attention heads	9 (3 KV heads)
Max sequence length	8192
Vocabulary size	36

Tokenizer

This model uses the REINVENT4 tokenizer — a chemistry-aware tokenizer that splits SMILES strings based on a hand-crafted regex covering atoms, bonds, ring closures, branches, and bracket atoms. The vocabulary has 36 tokens.

Usage

Pass an empty string to prompt the model to generate novel SMILES from scratch:

from transformers import AutoModelForCausalLM, PreTrainedTokenizerFast

model = AutoModelForCausalLM.from_pretrained("ddidacus/smolgen-pubchem-46M-base")
tokenizer = PreTrainedTokenizerFast.from_pretrained("ddidacus/smolgen-pubchem-46M-base")

inputs = tokenizer("", return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=1.0,
    num_return_sequences=10,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)

smiles_list = tokenizer.batch_decode(outputs, skip_special_tokens=True)
print(smiles_list)