Dataset Viewer
Auto-converted to Parquet Duplicate
table_id
stringlengths
10
10
family
stringlengths
2
17
n
int64
500
500
n_covariates
int64
0
4
truth
float64
-2
2
archive_path
stringlengths
27
27
train_0000
abadie_lhour
500
4
-1
train/tables/train_0000.npz
train_0001
santanna_zhao
500
4
-1
train/tables/train_0001.npz
train_0002
shi_iid
500
0
-0.837536
train/tables/train_0002.npz
train_0003
truncated
500
0
-0.029686
train/tables/train_0003.npz
train_0004
cwz_iid
500
0
0.571232
train/tables/train_0004.npz
train_0005
cwz_iid
500
0
1.382486
train/tables/train_0005.npz
train_0006
truncated
500
0
0.049612
train/tables/train_0006.npz
train_0007
truncated
500
0
0.038278
train/tables/train_0007.npz
train_0008
qin
500
2
-0.785974
train/tables/train_0008.npz
train_0009
zhou_network
500
0
-0.048605
train/tables/train_0009.npz
train_0010
abadie_lhour
500
4
-2
train/tables/train_0010.npz
train_0011
karmakar_censored
500
1
0.951124
train/tables/train_0011.npz
train_0012
shi_iid
500
0
-1.309067
train/tables/train_0012.npz
train_0013
zhao_bridge_panel
500
3
-0.698481
train/tables/train_0013.npz
train_0014
zhou_network
500
0
0.093462
train/tables/train_0014.npz
train_0015
zhao_bridge_panel
500
3
-0.518233
train/tables/train_0015.npz
train_0016
zhou_network
500
0
-0.066193
train/tables/train_0016.npz
train_0017
zhao_bridge_panel
500
3
1.978167
train/tables/train_0017.npz
train_0018
race_iid
500
1
-1.698052
train/tables/train_0018.npz
train_0019
zhou_network
500
0
-0.290586
train/tables/train_0019.npz
train_0020
race_iid
500
1
0.493896
train/tables/train_0020.npz
train_0021
zhou_network
500
0
-0.379598
train/tables/train_0021.npz
train_0022
qin
500
2
1.593231
train/tables/train_0022.npz
train_0023
qin
500
2
1.794961
train/tables/train_0023.npz
train_0024
zhou_network
500
0
-0.220989
train/tables/train_0024.npz
train_0025
ye
500
0
-0.481639
train/tables/train_0025.npz
train_0026
race_iid
500
1
0.135983
train/tables/train_0026.npz
train_0027
santanna_zhao
500
4
0
train/tables/train_0027.npz
train_0028
zhou_network
500
0
0.097074
train/tables/train_0028.npz
train_0029
abadie_lhour
500
4
2
train/tables/train_0029.npz
train_0030
qin
500
2
-0.695992
train/tables/train_0030.npz
train_0031
santanna_zhao
500
4
-2
train/tables/train_0031.npz
train_0032
karmakar_censored
500
1
-1.168504
train/tables/train_0032.npz
train_0033
mixture
500
0
0.016236
train/tables/train_0033.npz
train_0034
zhao_bridge_panel
500
3
1.323774
train/tables/train_0034.npz
train_0035
ferman_iid
500
1
1.756041
train/tables/train_0035.npz
train_0036
qin
500
2
0.969709
train/tables/train_0036.npz
train_0037
zhou_network
500
0
0.207758
train/tables/train_0037.npz
train_0038
qin
500
2
-0.37094
train/tables/train_0038.npz
train_0039
race_iid
500
1
-1.181091
train/tables/train_0039.npz
train_0040
cwz_iid
500
0
-1.421855
train/tables/train_0040.npz
train_0041
mixture
500
0
-0.423435
train/tables/train_0041.npz
train_0042
ye
500
0
1.731366
train/tables/train_0042.npz
train_0043
karmakar_censored
500
1
1.035834
train/tables/train_0043.npz
train_0044
mixture
500
0
-0.391251
train/tables/train_0044.npz
train_0045
ferman_iid
500
1
-0.029013
train/tables/train_0045.npz
train_0046
karmakar_censored
500
1
1.043078
train/tables/train_0046.npz
train_0047
ye
500
0
0.173222
train/tables/train_0047.npz

Statistical Inference Arena · Difference-in-Differences

Estimate the average treatment effect on the treated (ATT) in a two-period panel and return a two-sided 95% confidence interval. Each study contains 500 units and contributes one estimate. The benchmark compares a frozen PFN-DiD checkpoint and a cross-fitted random forest on the same 2,000 studies. Model

Purpose Studies Download
Historical adapter fitting 10,000 Training archive
Historical checkpoint selection 200 Validation archive
Try estimation and scoring 48 train, 16 dev, 32 test Starter kit
Evaluate a fixed method 2,000 Observed inputs, folds and prediction template
Reproduce the comparison Same 2,000 Population ATT labels and scorer

Target and observations

For each study, hold its data-generating mechanism and parameters fixed. The target is the average post-period effect of treatment among units that receive treatment.

ATT=E ⁣[Y1(1)−Y1(0)∣D=1]. \mathrm{ATT} = \mathbb{E}\!\left[Y_1(1)-Y_1(0)\mid D=1\right].

Here Y₁(1) and Y₁(0) are the same unit's potential outcomes in the post period under treatment and under control. The expectation averages over treated units in the population that generated the study. Report one estimate in outcome units and one two-sided 95% confidence interval per study.

Each NPZ in the evaluation archive holds baseline covariates X, a binary treatment indicator D, and the post-minus-pre outcome change delta_y, in matching row order. An array with shape (500, 0) denotes a study with no observed covariates. Supply only these three arrays to an estimator.

Run a first comparison

Extract the starter kit, install NumPy and run the following from the extracted directory.

python baseline.py dev/tables > predictions.csv
python score.py dev/truth.csv predictions.csv

The example computes the treated-minus-control difference in mean changes and uses a normal interval. Designs that require covariate adjustment can bias this simple estimate. The starter draws 96 separate studies from the 13-family prior with seed 20261119. Its three viewer splits describe the files inside the starter ZIP.

Historical data and evaluation design

The historical files contain the original adapter training and validation studies from ten families. Each archive preserves the original NPZs and manifest, and provides a table-ID-to-ATT CSV. The NPZs also contain simulation labels and metadata. Use only the observed arrays as model inputs.

The evaluation uses one pooled split. For each study the generator draws one of 13 families uniformly, draws its parameters, then samples 500 units. Seed 20261116 selects the exact tables used for the published comparison. The input archive preserves the observed NPZ bytes and row order. The viewer's evaluation_inputs configuration lists one row per study. Its SHA-256 column identifies the original input file. DGP families and realized counts

Evaluate a method

Freeze the estimator before opening the scoring archive. The evaluation labels are public for reproducibility. Keep the supplied five treatment-stratified folds for the original fit. Row indices start at 0 and fold labels run from 1 to 5. Write one row per table using the template columns below.

table_id,estimate,ci_lower,ci_upper

Run python score.py truth.csv predictions.csv after extracting the scoring archive. Each study has equal weight. Bias and RMSE use estimate minus population ATT. Error variance uses the sample variance of those errors (denominator 1,999). Coverage counts intervals containing the truth, including either endpoint. Mean interval length averages upper minus lower. Measure elapsed seconds per study during estimation and record the hardware if reporting runtime.

PFN-DiD and the RF comparator estimate the treatment propensity and the control-group mean outcome change, then apply the same doubly robust DiD score. PFN-DiD uses an influence-function interval. RF uses 50 whole-row bootstrap resamples with refitting and a normal interval based on the bootstrap standard deviation.

Downloads last month
71