Datasets:

csamplawski
/

SymbolicRegressionQA

dataset_info:
  features:
    - name: image
      dtype: image
    - name: choices
      large_list: string
    - name: label
      dtype: string
    - name: sympy_reprs
      large_list: string
    - name: degrees
      large_list: int32
    - name: question_id
      dtype: int64
  splits:
    - name: train
      num_bytes: 57268567
      num_examples: 10000
    - name: validation
      num_bytes: 5700650
      num_examples: 1000
    - name: test
      num_bytes: 5747661
      num_examples: 1000
  download_size: 57729824
  dataset_size: 68716878
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: cc-by-4.0

SymbolicRegressionQA

SymbolicRegressionQA (SRQA) is a multiple-choice visual question answering dataset introduced in Bayesian Adaptation Gym (UAI 2026). Each question shows a plot of a symbolic expression and asks which of four candidate formulas describes the relationship between the variables.

Splits

Split	Examples
train	10,000
validation	1,000
test	1,000

Fields

Field	Type	Description
`image`	`Image`	224x224 grayscale plot of the ground truth expression
`choices`	`list[str]`	Four candidate formulas, in the order they are presented
`label`	`str`	Letter of the correct choice, one of `A`, `B`, `C`, `D`
`sympy_reprs`	`list[str]`	SymPy representation of each of the four choices
`degrees`	`list[int]`	Polynomial degree of each of the four choices
`question_id`	`int`	Unique identifier within the split

The correct answer is close to uniform over the four letters in every split.

Usage

from datasets import load_dataset

data = load_dataset("csamplawski/SymbolicRegressionQA", split="test")

The prompt used in the BAG experiments is built from choices as follows:

For the provided image of a plot, which of following formulas best describes the relationship between the variables? Output the letter of your choice only.
Choices:
A) <choices[0]>
B) <choices[1]>
C) <choices[2]>
D) <choices[3]>

See bayesadapt/datasets/srqa.py in BayesAdapt for the loader used in the paper.

Citation

@misc{samplawski2026bayesianadaptationgymbenchmark,
      title={Bayesian Adaptation Gym: A Benchmark for the Bayesian Low-Rank Adaptation of Multi-Modal Language Models}, 
      author={Colin Samplawski and Ramneet Kaur and Manoj Acharya and Anirban Roy and Adam D. Cobb},
      year={2026},
      eprint={2606.22188},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.22188}, 
}