README.md

metadata

license: apache-2.0
task_categories:
  - graph-ml
tags:
  - biology
  - protein
  - molecule
  - dna
  - rna
  - pretraining

Cuttlefish-Encoder-Data

This repository contains the encoder pretraining dataset for Cuttlefish, a unified all-atom LLM designed for structure-grounded reasoning. It consists of all-atom structural graphs for molecules, proteins, DNA, and RNA in a unified parquet format, used for masked-reconstruction pretraining of the graph encoder.

• Paper: Scaling-Aware Adapter for Structure-Grounded LLM Reasoning
• GitHub Repository: zihao-jing/Cuttlefish
• Encoder Model: zihaojing/Cuttlefish-Encoder

Dataset structure

molecules/   # molecule encoder training data
nacid/       # DNA and RNA encoder training data
protein/     # protein encoder training data (PDB structures)
test/        # held-out test samples

Schema

Field	Description
modality	"molecule", "protein", "dna", or "rna"
node_feat	Atom/node features (N × d)
pos	3D coordinates in Å (N × 3)
edge_index	Spatial graph edges in COO (2 × E)
edge_feat_dist	Edge distances (E × 1, optional)

Usage

from datasets import load_dataset
ds = load_dataset("zihaojing/Cuttlefish-Encoder-Data")

Related resources

Resource	Link
Cuttlefish LLM	zihaojing/Cuttlefish
Cuttlefish-Encoder	zihaojing/Cuttlefish-Encoder
SFT instruction data	zihaojing/Cuttlefish-SFT-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 and Wang, Boyu and Hu, Pingzhao},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2602.02780}
}