README.md

metadata

language:
  - en
license: cc-by-4.0
size_categories:
  - n<1K
task_categories:
  - other
pretty_name: Stargazer
tags:
  - astronomy
  - exoplanets
  - benchmark
  - agents
  - radial-velocity
  - time-series
  - scientific-reasoning
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/default/train.jsonl
  - config_name: synthetic
    data_files:
      - split: train
        path: data/synthetic/train.jsonl
  - config_name: real
    data_files:
      - split: train
        path: data/real/train.jsonl

Dataset Card for Stargazer

Stargazer is a scalable environment for evaluating AI agents on dynamic, iterative physics-grounded model-fitting tasks using inference on radial-velocity (RV) time series data.

• Paper: Stargazer: A Scalable Model-Fitting Benchmark Environment for AI Agents under Astrophysical Constraints
• Project Page: https://aips-uoft.github.io/Stargazer/
• Repository: https://github.com/AIPS-UofT/Stargazer

Overview

Stargazer comprises 120 tasks across three difficulty tiers, including 20 real archival cases, covering diverse scenarios ranging from high-SNR single-planet systems to complex multi-planetary configurations requiring involved low-SNR analysis.

• Total tasks: 120
• License: cc-by-4.0

This release includes benchmark tasks built from both generated systems and archival RV systems, presented as one benchmark collection. Each row is one benchmark task with the original task payload preserved.

Usage

To run a single task by ID using the official repository:

python run_agent_batch_hard_timeout.py \
    --model gpt-5-mini \
    --task-ids seed22_diff4 \
    --verbose \
    --output-dir results_debug

To run all difficulty-1 tasks with 10 parallel workers:

python run_agent_batch_hard_timeout.py \
    --model gpt-5-mini \
    --count 1000 \
    --difficulties 1 \
    --workers 10 \
    --output-dir results_diff1

What Each Row Contains

Each row keeps the original benchmark structure:

• config
• observations
• difficulty_details
• meta

It also adds a few summary fields for filtering in the Hugging Face viewer:

• task_id
• truth_difficulty
• num_observations
• num_instruments
• num_planets
• engine
• has_gp_noise

Notes

• Tasks in this release may contain up to 7 planets
• This is a benchmark release, not a training corpus
• Ground-truth configuration fields are included, so they should not be exposed to the evaluated agent

Citation

@article{liu2026stargazer,
  title   = {Stargazer: A Scalable Model-Fitting Benchmark Environment for AI Agents under Astrophysical Constraints},
  author  = {Liu, Xinge and Zhang, Terry Jingchen and Sch{\"o}lkopf, Bernhard and Jin, Zhijing and Menou, Kristen},
  year    = {2026},
  journal = {arXiv preprint arXiv:2604.15664},
  url     = {https://arxiv.org/abs/2604.15664}
}