eltokh7/stata-econ-bench

Stata Econometrics Benchmark

250 natural-language econometrics & statistics tasks, each solvable with a short Stata program and graded by executing the generated code against 1250 hidden numeric test cases (5 per problem). This is an execution-based benchmark: a solution is correct only if running it reproduces the expected numeric result within a per-case tolerance — not by string match.

At a glance

• 250 problems, 1250 test cases (5 per problem)
• Target language: Stata
• Grading: execution-based, numeric comparison within per-case tolerance
• Difficulty tiers: 109 dev (development/visible) + 141 sacred (held-out)
• Topics: panel/FE regression, IV/2SLS, MLE, time series (VAR/IRF, unit roots), survival, matrix algebra, inequality & concentration indices, diagnostics (VIF, Durbin–Watson, Hausman, Chow), and more.

Schema

One JSON object per line. Problem-level fields:

field	type	description
id	string	Unique problem id (e.g. p2_001).
prompt	string	The task, in natural language. The model must return Stata code that solves it.
test_cases	list	5 independent graded cases (see below).
tier	string	dev or sacred.
language	string	Always stata.

Each entry in test_cases:

field	type	description
id	string	Case id (case_01 … case_05).
setup_code	string	Stata code that prepares the data/state (e.g. input … end, xtset). Runs before the model's solution.
expected_outputs	object	{ "type": "scalar", "key": <stata expression>, "value": <number> }.
tolerance	float	Max allowed absolute difference between the evaluated key and value.
language	string	Always stata.

Example

{
  "id": "p2_001",
  "prompt": "Run a fixed effects panel regression of y on x with entity fixed effects using xtreg. The data has panel structure with 'id' as the entity and 'time' as the time variable. Return the coefficient on x.",
  "test_cases": [
    {
      "id": "case_01",
      "setup_code": "clear\ninput id time x y\n1 1 1 5\n1 2 2 7\n1 3 3 9\n2 1 2 8\n2 2 3 10\n2 3 4 12\n3 1 1 6\n3 2 2 8\n3 3 3 10\nend\nxtset id time",
      "expected_outputs": {
        "type": "scalar",
        "key": "_b[x]",
        "value": 2.0
      },
      "tolerance": 1e-06,
      "language": "stata"
    },
    {
      "id": "case_02",
      "setup_code": "clear\ninput id time x y\n1 1 0 10\n1 2 1 13\n1 3 2 16\n2 1 1 5\n2 2 2 8\n2 3 3 11\nend\nxtset id time",
      "expected_outputs": {
        "type": "scalar",
        "key": "_b[x]",
        "value": 3.0
      },
      "tolerance": 1e-06,
      "language": "stata"
    }
  ],
  "tier": "sacred",
  "language": "stata"
}

How grading works

For each test case:

1. Run setup_code (loads/constructs the dataset, sets the panel/time structure, etc.).
2. Append and run the model's generated Stata solution.
3. Evaluate the Stata expression in expected_outputs.key and compare it numerically to expected_outputs.value.
4. The case passes iff abs(evaluated - value) <= tolerance.

A problem is typically scored as solved only when all 5 cases pass, which guards against solutions that hard-code one answer or overfit a single dataset.

Answer-key families

key is a Stata expression evaluated after the solution runs. The common families:

family	meaning	example
_b[...], _se[...]	estimated coefficient / standard error from the last estimation	_b[x]
r(...)	r-class returned results (e.g. summarize, correlate)	r(mean)
e(...)	e-class estimation results	e(N), e(r2)
result, named scalars	a problem-specific scalar the solution must compute and store	gini, hhi, F_chow, late

Observed key families in this release: r (435), _b (180), result (80), _se (25), e (20), hr_x (10), late (10), max_eigen (10), det_A (10), b_x (10), elem (10), hhi (5).

Tiers

• dev (109) — intended for development, prompt iteration, and visible evaluation.
• sacred (141) — held-out set kept separate from model development to give a cleaner generalization estimate.

Contamination note. This is a public release, so the sacred split is now on the open web and may eventually appear in pretraining corpora. Treat scores accordingly and prefer fresh/private problems for high-stakes leaderboard claims.

Usage

from datasets import load_dataset

ds = load_dataset("eltokh7/stata-econ-bench", split="test")
print(ds[0]["prompt"])
print(ds[0]["test_cases"][0]["expected_outputs"])  # {'type': 'scalar', 'key': '_b[x]', 'value': 2.0}

# Filter to held-out problems
sacred = ds.filter(lambda r: r["tier"] == "sacred")

Or read the raw JSONL directly:

import json
data = [json.loads(line) for line in open("stata_econ_bench.jsonl")]

You need a working Stata installation (batch mode) to actually grade generated solutions.

License

Released under CC-BY-4.0. If you use this benchmark, please credit the source.

@misc{stata_econ_bench,
  title  = {Stata Econometrics Benchmark},
  author = {Eltoukhy, Khaled},
  year   = {2026},
  howpublished = {Hugging Face Datasets},
  url    = {https://huggingface.co/datasets/eltokh7/stata-econ-bench}
}