NicolasNoya/molecule-gat-contrastive

Molecule GAT - Contrastive Learning

This model was trained for the ALTEGRAD Challenge 2025: Molecule-to-Text retrieval.

Model Description

• Architecture: GAT (Graph Attention Network)
• Task: Learn molecular graph embeddings for text retrieval
• Training: Contrastive learning (CLIP-style InfoNCE loss)
• Embedding Dimension: 768

Architecture Details

GAT(
    input_dim=N/A,
    hidden_dim=512,
    out_dim=768,
    num_layers=3,
    num_heads=4,
    dropout=0.1,
    pooling='mean'
)

Training Details

• Loss: InfoNCE (Contrastive Loss)
• Temperature: 0.07
• Best Validation MRR: N/A
• Dataset: ALTEGRAD Challenge 2025

Usage

import torch
from torch_geometric.data import Data
from huggingface_hub import hf_hub_download

# Download model files
checkpoint_path = hf_hub_download(
    repo_id="NicolasNoya/molecule-gat-contrastive",
    filename="model.pt"
)

# Load checkpoint
checkpoint = torch.load(checkpoint_path, map_location='cpu')

# Reconstruct model (you'll need the model class definition)
# See model_architecture.py in this repo

model = MolGAT(
    input_dim=N/A,
    hidden_dim=512,
    out_dim=768,
    num_layers=3,
    num_heads=4,
    dropout=0.1,
    pooling='mean'
)

model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Use the model
with torch.no_grad():
    # graph_data should be a PyTorch Geometric Data object
    embedding = model(graph_data)

Model Performance

Evaluated on ALTEGRAD Challenge 2025 validation set:

• MRR (Mean Reciprocal Rank): N/A
• Recall@1: N/A
• Recall@5: N/A
• Recall@10: N/A

Citation

If you use this model, please cite:

@misc{molecule_gat_2025,
  author = {NicolasNoya},
  title = {Molecule GAT for Contrastive Learning},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/NicolasNoya/molecule-gat-contrastive}}
}

License

MIT License

Contact

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