jonghyunlee/MolGPT_long_context_pretrained-by-ZINC15

🧪 Mol-GPT2; long context, pretrained with ZINC-15

This repository hosts a GPT-2-based model for generating SMILES strings, trained on the ZINC 15 dataset. The model follows the architecture and hyperparameter setup of MolGPT (Bagal et al., 2021), and has been fine-tuned to generate valid molecular representations with high accuracy. This model has longer context length (256), whereas the previous model has maximum context lenght 128.

---

🔧 Model Architecture

GPT2Config( vocab_size=10_000, n_positions=256, n_ctx=256, n_embd=256, n_layer=8, n_head=8, resid_pdrop=0.1, embd_pdrop=0.1, attn_pdrop=0.1, )

• Pretrained with fp16 precision on 2× H100 GPUs
• Batch size: 1,024
• Max steps: 100,000
• Warmup steps: 10,000
• Evaluation every 10,000 steps

---

📊 Performance

Dataset/Metric	This Model	Short Context	MolGPT Baseline
ZINC15 Validity	99.76%	99.68%	N/A
MOSES Validity	N/A	N/A	99.4%
GuacaMol Validity	N/A	N/A	98.1%

---

🚀 Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("jonghyunlee/MolGPT_long_context_pretrained-by-ZINC15")
model = AutoModelForCausalLM.from_pretrained("jonghyunlee/MolGPT_long_context_pretrained-by-ZINC15", torch_dtype=torch.float16)

# Generate molecules
input_ids = tokenizer("CC(=O)OC1=CC=CC=C1C(=O)O", return_tensors="pt").input_ids
outputs = model.generate(input_ids, max_length=256, do_sample=True, top_k=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))