pretty_name: FindStatBench
license: cc-by-sa-3.0
language:
- en
task_categories:
- text-generation
tags:
- code-generation
- mathematics
- combinatorics
- benchmark
size_categories:
- 1M<n<10M
configs:
- config_name: prompts
default: true
data_files:
- split: test
path: data/prompts/*.parquet
- config_name: tasks
data_files:
- split: test
path: data/tasks/*.parquet
- config_name: hidden
data_files:
- split: test
path: data/hidden/*.parquet
FindStatBench
FindStatBench evaluates whether a model can turn a mathematical description into a Python function, solve(obj), that computes a statistic or maps one combinatorial object to another. The corpus contains 2,329 tasks in 24 collections and 5,517,046 evaluation instances held out from generation prompts.
This Parquet distribution, version 1.0.0-hf1, preserves the original task records, closed-book prompts, and evaluation instances without changing their values or order. See the paper for the evaluation protocol and the Zenodo archive for the historical evaluator and experiment artifacts.
Configurations
| Configuration | Rows | Contents |
|---|---|---|
prompts (default) |
2,329 | Exact model-facing closed-book prompts and task IDs |
tasks |
2,329 | Task metadata, public examples, and original source-bearing task prompts |
hidden |
5,517,046 | Evaluation inputs and reference outputs, withheld from model prompts |
All configurations have a single test split. The corpus includes 1,993 statistic tasks (1,929,082 evaluation instances) and 336 map tasks (3,587,964 evaluation instances).
Loading the data
from datasets import load_dataset
repo_id = "sdan021/FindStatBench"
prompts = load_dataset(repo_id, "prompts", split="test")
evaluation = load_dataset(repo_id, "hidden", split="test", streaming=True)
For the one-shot protocol, send only the prompt field from the prompts configuration. The tasks.prompt field is a source-bearing archival prompt, not the input for closed-book generation. The evaluation labels are publicly downloadable; they must not be used for training, retrieval, feedback, or candidate selection when reporting closed-book scores. Report public-example repair and any other tool use or selection separately.
Nested fields in the task and evaluation tables are encoded as JSON strings in columns ending in _json. This preserves strings such as binary words with leading zeros. The MIT-licensed findstatbench_data.py helper decodes those fields without changing the records. Record the repository revision for reproducibility. package_manifest.json records source and shard checksums and canonical record digests.
Evaluation and limitations
Use the historical evaluator in the Zenodo archive for comparison with the paper. Run untrusted generated programs only in an isolated execution environment. An extracted program can fail to run or be incorrect. Report task accuracy (all cases correct for a task) and instance accuracy (fraction of cases correct) separately.
Some inputs and map targets retain normalized string encodings. Exact scoring can reject mathematically equivalent outputs in other representations. Public source pages, retained cross-references, and downloadable labels mean this benchmark cannot establish that a model has never seen a task. Finite tests are not a proof of correctness, and these results are not a measure of general software engineering.
Attribution and licensing
FindStatBench is maintained by Soham Dan and derives from FindStat, maintained by Martin Rubey, Christian Stump, and the FindStat contributors. Cite the FindStatBench paper and follow the FindStat citation guidance.
The data adaptations follow the upstream CC BY-SA 3.0 Unported policy to the extent of the applicable upstream grant. See attribution and changes, data rights, and the full license. The original decoding helper is separately MIT-licensed. Model-provider outputs and generated programs are not included here.