Views
No views yet
| Property | Value |
|---|---|
| Architecture | H-Net (Hierarchical Network) |
| Parameters | ~350M |
| Dataset | MOSES |
| Training Bytes | 340M |
| Training Epochs | 5 |
| Concatenation | 10x SMILES per example |
| Architecture Variant | 2-stage |
['m4', ['T1m4', ['T22'], 'm4T1'], 'm4']checkpoints/checkpoint_bytes_best.pt - Best checkpoint (lowest validation loss)checkpoints/checkpoint_epoch_*.pt - Epoch checkpointsmetadata.json - Training configuration and historytest_smiles.txt - Test SMILES used during trainingvisualizations/ - Training evolution GIFs and prediction files1import torch
2from pathlib import Path
3
4# Load checkpoint
5checkpoint_path = "checkpoints/checkpoint_bytes_best.pt"
6checkpoint = torch.load(checkpoint_path, map_location="cpu")
7
8# The checkpoint contains:
9# - 'model_state_dict': Model weights
10# - 'optimizer_state_dict': Optimizer state
11# - 'epoch': Training epoch
12# - 'metrics': Training metrics
13# - 'cumulative_training_bytes': Total bytes processed
14
15# Load into your H-Net model
16# model.load_state_dict(checkpoint['model_state_dict'])| Metric | Value |
|---|---|
| Bits-per-byte (BPB) | 0.83 |
| Mean token length | 2.0 |
1@inproceedings{hnet_smiles_2026,
2 title={Learning Chemical Grammar: Dynamic Tokenization for SMILES with Hierarchical Networks},
3 author={Anonymous},
4 booktitle={International Conference on Machine Learning (ICML)},
5 year={2026}
6}