What the agent receives
Target and explanatory variables, controls, estimator, analysis requirements, and mounted source data.
DATA · SELECTED_1000
The benchmark, task by task.
Each task names an outcome, a treatment, controls, an estimation method and a data file. The goal is to recover one reported effect: its coefficient, standard error and p-value.
Dataset downloads require a Hugging Face account with access. Open the dataset to request access before using the download link.
1,000
fixed evaluation tasks
15
journals in this subset
286
distinct article titles
5
estimation classes
Selected_1000 · Counts from the pinned task list, revision 59f9512.
01 / DISTRIBUTION
Different research questions. Different demands.
Estimation methods
Five classes · 1,000 tasks
OLS includes panel OLS; IV includes IV-2SLS. Other combines Probit, Logit, negative binomial, LMM, PSM and Cox models.
Journal coverage
The five largest sources; all 15 below
All 15 journals
Percentages use all 1,000 tasks. These distributions are recomputed from the pinned CSV, not from the moving dataset main branch.
02 / TASK FORMAT
A clear input and output contract.
What the agent produces
An executable program and a result JSON containing coefficient, standard_error and p_value.
What the evaluator keeps separate
Reference answers and replication programs are kept outside the model input during evaluation. Dataset downloads can contain references.
03 / EXAMPLE
One task, in context.
TASK 0011 · OLS
Economic uncertainty and the “two Spains”
Recover the relationship between socio-economic conflict and policy uncertainty, with three controls and month-clustered standard errors.
See the agent walkthrough →- y
-
EPU0month_simsn_w - x
-
Wscmonth_simsn_w - Data
-
data_np.dta - Sample
- 1905–1945 · 977 estimation observations
04 / SOURCE
A fixed, inspectable snapshot.
- Dataset
- CamoAiLab/InferenceNet ↗
- Pinned revision
-
59f9512a38e594528807744214a60ee00367434e - Task list
-
Selected_1000/1000_new.csv - Subset labels
- 1,000 Stata-labelled tasks; 14 raw method labels folded into five classes. Article count is the number of distinct title strings.