metadata
license: apache-2.0
pipeline_tag: graph-ml
tags:
- biology
- protein
- molecule
- dna
- rna
- graph-neural-network
Cuttlefish-Encoder
Graph encoder component of Cuttlefish, a unified all-atom LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity.
This model was presented in the paper Scaling-Aware Adapter for Structure-Grounded LLM Reasoning.
- Code: GitHub - zihao-jing/Cuttlefish
- Pretrained with: Masked reconstruction on all-atom structures.
Usage
You can download the encoder using the huggingface_hub library:
from huggingface_hub import snapshot_download
encoder_dir = snapshot_download("zihaojing/Cuttlefish-Encoder")
# Load via the Cuttlefish codebase
# See https://github.com/zihao-jing/Cuttlefish for full usage
Pretraining data
Pretrained on Cuttlefish-Encoder-Data, covering:
- Molecules (SMILES → 3D graph)
- Proteins (PDB/CIF → all-atom graph)
- DNA and RNA sequences
Model details
- Architecture: All-atom graph encoder with Scaling-Aware Patching.
- Encoder hidden dim: 256
- Modalities: molecule, protein, dna, rna
Related resources
| Resource | Link |
|---|---|
| Full Cuttlefish LLM | zihaojing/Cuttlefish |
| SFT instruction data | zihaojing/Cuttlefish-SFT-Data |
| Encoder pretraining data | zihaojing/Cuttlefish-Encoder-Data |
Citation
@article{jing2026cuttlefish,
title = {Cuttlefish: Scaling-Aware Adapter for Structure-Grounded LLM Reasoning},
author = {Jing, Zihao and Zeng, Qiuhao and Fang, Ruiyi and Li, Yan Yi and Sun, Yan Table, Boyu and Hu, Pingzhao},
booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
year = {2026},
url = {https://arxiv.org/abs/2602.02780}
}