Curie Dataset

HF version of the dataset:

CURIE: Evaluating LLMs On Multitask Scientific Long Context Understanding and Reasoning.

Also available via GitHub (Apache-2.0 license).

Dataset Structure

CURIE consists of 10 tasks that are mapped to 8 datasets. The datasets are:

Dataset ID	Task Name	Domain	Description
biogr	Biodiversity Georeferencing	Biodiversity	Determine the latitude, longitude bounding box encompassing the region in the map image.
dft	Density Functional Theory Analysis	Condensed Matter Physics	3 Tasks related to DFT.
pdb	Protein Sequence Reconstruction	Protein Sequencing	Reconstruct a protein’s amino acid sequence form the 3D structure.
geo	Geospatial Dataset Extraction	Geospatial Analysis	Extract information for all geospatial datasets used along with the spatial and temporal extents.
mpve	Materials Property Value Extraction	Materials Science	Identify all instances of materials,their properties, and descriptors
qecc	Quantum Error Correction Codes	Quantum Computing	Create a YAML file with the Error Correction Code’s properties.
hfd	Hartree-Fock Tasks Derivation	Condensed Matter Physics	Derive the Hartree-Fock mean-field Hamiltonian for a quantum many-body system
hfe	Hartree-Fock Tasks Extraction	Condensed Matter Physics	Extract the most general mean-field Hamiltonian.

Each dataset contains the fields:

id (str): The sample id.
prompt (str): The prompt containing the task description for the LLM.
text (str): The sample specific information needed to solve the task.
gt (str): The groundtruth answer as a json-string. To obtain the structured representation load the string with json5 (see Example).
difficulty_level (str): Difficulty level of the task.

Special fields for some datasets:

Dataset	Additional Fields	Content
biogr	figure (PIL Image)	Figure containing geographical map
dft	prompt_metadata (str), prompt_structure_data (str)	Prompts for the subtasks dft-structure & dft-metadata
mpve	prompt_exclude_trivia (str) prompt_bandgap_refractive (str)	Ablation prompts

Example Usage

Example: Query gpt-4o for a response on the hfd dataset using a LangChain chat model:

import json5

from datasets import load_dataset
from langchain.chat_models.base import init_chat_model

dataset = load_dataset('nhop/curie','hfd')
llm = init_chat_model('gpt-4o')

for sample in dataset["train"]:
  print(sample["prompt"])
  prompt = sample["prompt"].replace("{{text}}",sample["text"])
  response = llm.invoke(prompt)
  print(response.content)
  groundtruth = json5.loads(sample["gt"])
  print(groundtruth)
  break

Citation

@inproceedings{cui2025curie,
  title={CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning},
  author={Cui, Hao and Shamsi, Zahra and Cheon, Gowoon and Ma, Xuejian and Li, Shutong and Tikhanovskaya, Maria and Norgaard, Peter Christian and Mudur, Nayantara and Plomecka, Martyna Beata and Raccuglia, Paul and others},
  booktitle={The Thirteenth International Conference on Learning Representations}
  year={2025}
}