wytzzka commited on
Commit
2027f34
·
verified ·
1 Parent(s): 5850dab

Upload 11 files

Browse files
DGP.md ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Proximal starter v1: mechanisms and parameter priors
2
+
3
+ Each table is a 500-observation study and is scored at one independent baseline
4
+ query x*. Target: `CATE(x*) = E[Y(1)-Y(0)|X=x*]`.
5
+
6
+ ## Common proxy model
7
+
8
+ Let `X1,X2 iid N(mu, sigma_x^2)`, `c=(.25,.25)` and
9
+ `p(x)=1/[1+exp{.125*r_p*(x1+x2)}]`. The observation sampler draws
10
+ `A|X ~ Bernoulli(p(X))`, then jointly Gaussian `(Z,W,U)|A,X` with mean
11
+
12
+ `(.25,.25,.25) + A*(.25,.25*r_w,.25) + (c'X)*(1,1,1)`
13
+
14
+ and covariance
15
+
16
+ ```
17
+ [[1, r_z*r_w, r_z],
18
+ [r_z*r_w, 1, r_w],
19
+ [r_z, r_w, 1 ]].
20
+ ```
21
+
22
+ This is a **factorization of a joint distribution**, not a claim that treatment
23
+ causally changes pre-treatment latent U or the baseline proxies. The coupled
24
+ covariance and W-shift preserve the intended Gaussian conditional proxy
25
+ restrictions. Let
26
+
27
+ `m_W(u,x)=.25+.25(x1+x2)+r_w*(u-.25-.25(x1+x2))`,
28
+
29
+ `g_W(x)=E[W|X=x]=.25+.25(x1+x2)+.25*r_w*p(x)`.
30
+
31
+ Generate
32
+
33
+ `Y = 2 + f(X) + beta_W*m_W(U,X) + 2*(W-m_W(U,X)) + s*A*delta(X,m_W(U,X)) + epsilon`.
34
+
35
+ The result bridge is `h(W,A,X)=2+f(X)+beta_W*W+s*A*delta(X,W)`.
36
+ Because each delta is affine or constant in W, the analytic query effect is
37
+ `tau(x)=s*delta(x,g_W(x))`. Natural W|X is integrated out; W|A,X is not substituted.
38
+ U and references never enter the observed table.
39
+
40
+ ## Four families
41
+
42
+ | Family | f(x) | beta_W | delta(x,w) | Noise |
43
+ |---|---|---|---|---|
44
+ | Cui linear | .25(x1+x2) | 4q | 2 | N(0,(.25v)^2) |
45
+ | Qi heterogeneous linear | .25(x1+x2) | 8q | .5+3x1-5x2+c4*w | Uniform(-v,v) |
46
+ | Qi quadratic | x1^2+x2^2 | 5q | .25-6x1*x2 | Uniform(-v,v) |
47
+ | Qi trigonometric | x1^2+x2^2 | 4q | 2.3+abs(x1-1)-abs(x2+1)+(-2.5+sin(x1)-2cos(x2))*w | Uniform(-v,v) |
48
+
49
+ Qi linear first samples L1/L2 equally, with c4=.25 or 0. These scenarios are
50
+ not counted as extra independent mechanism families. Query x* is a fresh
51
+ independent draw from the same `N_2(mu*1, sigma_x^2 I)` covariate population.
52
+
53
+ ## Hyperprior and source adaptations
54
+
55
+ Family probabilities are 1/4 each. Within a task the independent continuous draws
56
+ are `s~U(.5,1.5)`, `v~LogUniform(.5,2)`, `mu~U(0,.5)`,
57
+ `sigma_x~U(.2,.4)`, `r_p~U(.5,1.5)`, `q~U(.75,1.25)`, and
58
+ `r_z,r_w~U(.35,.7)`. Covariance and W mean-shift are then updated jointly.
59
+ These are benchmark hyperpriors, not a claim of exact source-paper settings.
60
+
61
+ - Cui, Pu, Shi, Miao & Tchetgen Tchetgen (2024), *Semiparametric Proximal Causal Inference*, JASA 119(546),1348-1359, [DOI:10.1080/01621459.2023.2191817](https://doi.org/10.1080/01621459.2023.2191817). Mechanism locator: arXiv:2011.08411v4 §5.1, PDF21-22, equations13-14.
62
+ - Qi, Miao & Zhang (2024), *Proximal Learning for Individualized Treatment Regimes Under Unmeasured Confounding*, JASA119(546),915-928, [DOI:10.1080/01621459.2022.2147841](https://doi.org/10.1080/01621459.2022.2147841). Mechanism locators: §6.1 Table3 L1/L2; Table4 N2 (quadratic), N1 (trigonometric), and author ProxITR code.
63
+
64
+ Qi treatment coding is adapted to 0/1 and the target to a query CATE. Extraneous
65
+ noise covariates and auxiliary source labels are not imported. In particular,
66
+ the source N2 helper labels with an inconsistent branch call are not used;
67
+ the reference follows the displayed N2 outcome/bridge equations.
68
+
69
+ ## Public starter generation
70
+
71
+ The shipped starter uses 48 train,16 dev,32 test tables, each from a distinct
72
+ sampled parameter configuration. Each configuration contributes **one table**.
73
+ They have fresh seed namespaces and do not reproduce official observations.
74
+ The labels in each split's truth.csv are exact analytic query references.
75
+ This public seed is for the starter only and is not the official collection seed.
76
+
77
+ ```
78
+ python generate.py --output extra --seed 314159 --train 2000 --dev 16 --test 32
79
+ ```
80
+
81
+ Requires Python3.10+ and NumPy1.26 or later. The generator refuses to replace
82
+ an existing output directory. `catalog` has public labels, queries are stored
83
+ separately, and observations use round-trip float64 CSV precision. New training
84
+ draws should not replace the saved official inputs.
85
+
86
+ ## Official design and scoring boundary
87
+
88
+ The frozen official collection has 2,000 different parameter draws, one table
89
+ and one query each, without ID/OOD labels. It cannot estimate a within-fixed-
90
+ configuration sampling variance from paired repeats. Report across-task error
91
+ variance with N-1 denominator, with raw and optional SD(Y)-normalized scales
92
+ kept separate. Nominal 95% intervals do not guarantee coverage.
93
+
94
+ Public illustrative mechanisms, equations and technical consistency checks
95
+ do not replace scientific review or establish identification for every possible
96
+ adaptation. The source registry's `needs_scientific_review` state is retained.
README.md ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: Statistical Inference Arena - Proximal Causal Inference
3
+ language:
4
+ - en
5
+ tags:
6
+ - tabular
7
+ - synthetic
8
+ - causal-inference
9
+ - proximal-causal-inference
10
+ - benchmark
11
+ configs:
12
+ - config_name: starter_catalog
13
+ default: true
14
+ data_files:
15
+ - split: train
16
+ path: catalog/train.csv
17
+ - split: validation
18
+ path: catalog/dev.csv
19
+ - split: test
20
+ path: catalog/test.csv
21
+ - config_name: official_inputs
22
+ data_files:
23
+ - split: test
24
+ path: official_schema.csv
25
+ ---
26
+
27
+ # Statistical Inference Arena · Proximal Causal Inference
28
+
29
+ Given an observational study with **500 observations**, observed confounding proxies
30
+ and one independent covariate query, estimate the **conditional average treatment
31
+ effect** and return a two-sided 95% interval:
32
+
33
+ $$\tau(x^*)=E[Y(1)-Y(0)\mid X=x^*].$$
34
+
35
+ Both classical estimators and tabular foundation models can use the inputs.
36
+ [Arena leaderboard](https://statsinference.github.io/#domain=proximal) ·
37
+ [Official ProximalFM model](https://huggingface.co/christophemuller/ProximalFM) ·
38
+ [DGP equations and priors](DGP.md)
39
+
40
+ This is a **data repository**, following the official-input/starter separation in
41
+ the [mediation data release](https://huggingface.co/datasets/shibo769/StatsInference-Arena-Mediation).
42
+ It does not contain a new trained model or claim official Arena endorsement.
43
+
44
+ | What do you want to do? | Tables | Ground truth | Download |
45
+ |---|---:|---|---|
46
+ | Try observed-data loading, predictions and local scoring | 48 train, 16 dev, 32 test | Public | [Starter kit](proximal-starter-v1.zip) |
47
+ | Evaluate a fixed method on the recorded benchmark inputs | 2,000 | Separate maintainer scoring materials | [Official inputs and submission CSV](proximal-official-inputs-v1.zip) |
48
+
49
+ No historical ProximalFM pretraining, adapter-fitting or fine-tuning data are
50
+ redistributed. The starter's `train` split is a small **new public illustration**,
51
+ not data previously used to train the official model. The evaluated 2,000-table
52
+ collection has **no ID/OOD split**.
53
+
54
+ ## Target and observations
55
+
56
+ Every official `tables/<table_id>.csv` contains columns `X1,X2,Z,W,A,Y`:
57
+
58
+ | Columns | Meaning | Type |
59
+ |---|---|---|
60
+ | X1, X2 | Baseline covariates defining the CATE subgroup | Continuous |
61
+ | Z | Treatment-confounding proxy | Continuous |
62
+ | W | Outcome-confounding proxy | Continuous |
63
+ | A | Treatment | Binary 0/1 |
64
+ | Y | Outcome | Continuous |
65
+
66
+ `queries.csv` supplies one `X1,X2` query per `table_id`. Query Z/W/A/Y are not
67
+ observed or required. U is latent and is never included in the observed inputs.
68
+ This is a query-point population CATE, not ATT or a within-table average effect.
69
+ The target averages over the natural `W|X=x*` distribution under the stated
70
+ proximal assumptions. These assumptions require more than correlation between
71
+ the proxies and the hidden confounder.
72
+
73
+ ## Run the public starter
74
+
75
+ Extract `proximal-starter-v1.zip` and run from its directory:
76
+
77
+ ```sh
78
+ python -m pip install numpy
79
+ python baseline.py --tables dev/tables --queries dev/queries.csv --output dev_predictions.csv
80
+ python score.py --truth dev/truth.csv --predictions dev_predictions.csv
81
+ ```
82
+
83
+ The illustrative baseline is a linear S-learner on X/A, with treatment-by-X
84
+ interactions and B=50 row-bootstrap percentile intervals. It **ignores proxies**
85
+ and has no proximal identification guarantee. It is not the evaluated RF/GMM
86
+ comparator. On the 16 development tables its saved example has bias **0.637534**,
87
+ RMSE **1.024240**, coverage **62.5%**, and mean length **1.911591**. Those are
88
+ interface/scoring illustrations, not leaderboard results or tuned performance.
89
+
90
+ Replace that estimator, retain the four output columns and use the scorer. Tune
91
+ on `train`/`dev`; keep the public `test` split for a final local check. These
92
+ 96 fresh tables, their parameter draws and queries are separate from the frozen
93
+ 2,000 official tables. Each starter configuration contributes **one table**.
94
+
95
+ ## Sampling and file structure
96
+
97
+ **Public starter.** A family is sampled uniformly from Cui linear bridge, Qi
98
+ heterogeneous linear, Qi quadratic and Qi trigonometric; then source scenario
99
+ and continuous parameters are drawn, followed by 500 observations and an
100
+ independent query. Train/dev/test use distinct split-specific seed namespaces.
101
+ There is one new parameterized mechanism per table, not paired replicates.
102
+ `catalog/{train,dev,test}.csv` contains public labels and table paths within the
103
+ kit. The Dataset Viewer calls `dev` the `validation` split. A generator is
104
+ included for independent illustrations. [Equations and sources](DGP.md).
105
+
106
+ **Official inputs.** The collection preserves all 2,000 frozen tasks evaluated
107
+ on 2026-11-16, n=500. Family counts are 496, 516, 510 and 478 in the order above.
108
+ Each task has a separate continuously sampled configuration and query. Unlike
109
+ the mediation release's paired design, this collection does **not** contain
110
+ 1,000 configurations with two repeated tables. Its error variance is across
111
+ tasks, not within-configuration sampling variance.
112
+
113
+ The archive provides observed CSVs, queries, variable types, the saved original
114
+ five treatment-stratified folds and a blank prediction template. Opaque IDs
115
+ replace internal task indices. Row order, query values and original fold labels
116
+ are preserved. CSV uses round-trip float64 precision, without standardization,
117
+ trimming, rounding to float32 or changing the observations.
118
+
119
+ Direct table-level official true effects, generator assignments, parameters,
120
+ seeds, original IDs and reference model outputs are **not included in this public
121
+ input release**. They are retained in a separate maintainer package. This is a
122
+ release boundary, **not a cryptographic secrecy guarantee**: anyone with other
123
+ matching labeled artifacts or generation provenance may be able to link or
124
+ reconstruct tasks. Published aggregate results and analytic equations are not
125
+ withdrawn. Do not train or tune on references obtained elsewhere and claim an
126
+ unseen held-out evaluation.
127
+
128
+ ## Evaluate and return predictions
129
+
130
+ Freeze your method/settings before evaluating the official inputs. If your
131
+ method uses cross-fitting or context partitions, use
132
+ `folds/<table_id>.csv` aligned to the observed rows. Each contains one `fold`
133
+ column, labels 0-4, 100 observations per fold. Fold labels are evaluator
134
+ metadata, not features. Bootstrap refits must apply the declared resplitting
135
+ procedure; original labels are not prescribed as labels for every resample.
136
+
137
+ Fill `predictions_template.csv` with exactly one row per official table:
138
+
139
+ ```csv
140
+ table_id,estimate,ci_lower,ci_upper
141
+ ```
142
+
143
+ All numeric values must be finite and `ci_lower <= ci_upper`. Return effects
144
+ in **original outcome units**. The scoring reference uses the same query for
145
+ all methods on each task. Maintainers can compute signed Bias, sample error
146
+ variance with denominator 1,999, RMSE, 95% interval coverage and mean length.
147
+ They can separately standardize per-task error/length by the original table's
148
+ sample SD(Y); normalized and raw metrics must not be mixed. Runtime is not
149
+ inferred from the four-column file: submit hardware and warm point-plus-interval
150
+ timing, including all RF bootstrap work, separately when available.
151
+
152
+ Uploading this dataset does not submit a method or automatically update the
153
+ leaderboard. Share the resulting dataset URL with the Arena maintainers. No
154
+ private scoring files need to be uploaded or posted in a public repository.
155
+
156
+ ## Verification
157
+
158
+ Run `python verify_release.py` in a complete download of this repository. It
159
+ checks SHA256SUMS, both archives, all 2,000 input tables and folds, 96 labeled
160
+ starter tables and template IDs. No model or official truth is required.
161
+
162
+ The original formal evaluation inputs/results are unchanged; only public file
163
+ format, table IDs and release organization are adapted. No new official-model
164
+ evaluation or recalibration was performed for this data release.
165
+
166
+ ## References and scope
167
+
168
+ - Cui et al. (2024). *Semiparametric Proximal Causal Inference*. [DOI](https://doi.org/10.1080/01621459.2023.2191817), §5.1, equations (13)-(14).
169
+ - Qi et al. (2024). *Proximal Learning for Individualized Treatment Regimes Under Unmeasured Confounding*. [DOI](https://doi.org/10.1080/01621459.2022.2147841), §6.1, Tables 3-4.
170
+ - Muller et al. (2026). *ProximalFM: Amortized Proximal Causal Inference under Hidden Confounding*. [Paper](https://arxiv.org/abs/2610.08078v2) · [Official code](https://github.com/ChristopheMuller/ProximalFM).
171
+
172
+ Four mechanism families share a Gaussian proxy skeleton and are explicit
173
+ continuous-prior adaptations, not complete reproductions of all source tasks.
174
+ Original scientific-review status remains `needs_scientific_review`; finite
175
+ numerical and engineering checks do not prove universal identification or
176
+ interval calibration. No license for third-party models, papers or code is
177
+ inferred from the mediation example; their original terms remain applicable.
RELEASE_STATUS.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_date": "2026-11-19",
3
+ "release_kind": "Dataset public inputs and independent public starter",
4
+ "reference_layout": "shibo769/StatsInference-Arena-Mediation",
5
+ "reference_revision": "00fffb4e26e3d8be542c92319a81dec75cfa4d39",
6
+ "formal_tables": 2000,
7
+ "observations_per_table": 500,
8
+ "official_parameter_configurations": 2000,
9
+ "official_replicates_per_configuration": 1,
10
+ "starter_tables": {
11
+ "train": 48,
12
+ "dev": 16,
13
+ "test": 32
14
+ },
15
+ "starter_is_new_illustration": true,
16
+ "observations_queries_and_original_fold_labels_preserved": true,
17
+ "official_CSV_roundtrip_float64": true,
18
+ "formal_results_changed": false,
19
+ "formal_models_rerun": false,
20
+ "official_truth_in_public_package": false,
21
+ "model_weights_or_private_connection_material_in_public_package": false,
22
+ "automatic_upload_performed": false,
23
+ "privacy_scope": "Direct official references withheld here; no promise of non-reconstructability from other matching material.",
24
+ "validation": {
25
+ "public_structure_and_checksums_passed": true,
26
+ "all_official_CSV_values_roundtrip_equal_to_frozen_arrays": true,
27
+ "queries_and_original_folds_preserved": true,
28
+ "public_starter_references_independently_recomputed": 96,
29
+ "starter_overlaps_official_arrays": false,
30
+ "maintainer_reference_PFN_and_RF_metrics_match_formal_summary": true
31
+ }
32
+ }
SHA256SUMS ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ 24b254f9dcbde261f5010956a89874d87582c31efcf773df4499d4149ac89147 catalog/dev.csv
2
+ e872660afdb3e5de7ce78460e69b7b272ce621e47805419d848bdede7d3e53cf catalog/test.csv
3
+ a9d4a803519a4e30a97fc6b0ca2af641862d2863331f4de5e962844caeba3443 catalog/train.csv
4
+ f6665cb2a88fd1dd6dcfa0cc4393bb28eddd8996823cb29781cfbce01142538b DGP.md
5
+ 5f1e1de37750069e96562aae77d76a9b1988a57a0925be33f0072f308d963934 official_schema.csv
6
+ 01603b2c869715ec9486b9a0e73696b632d009a86c77a0f02d1dc69e387bafd6 proximal-official-inputs-v1.zip
7
+ 55bc8e29b91bdcc440a3888a28db57644cd415558836a74d5096c2b4956bb079 proximal-starter-v1.zip
8
+ beff5686179949896ee7e3e786de4a5f96e8a0d5773f215dc80e13fa14abec00 README.md
9
+ bbfb3c34df58cd6f513871c39d5c0dfcc093e878c6417b3617f6c66c109d01c4 RELEASE_STATUS.json
10
+ ffc73beeef4f4a8e45a8523fb1ce110ba5090a64a8d24e847f6c23adcf219589 verify_release.py
catalog/dev.csv ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ table_id,preset,config_id,replicate,n,truth,archive_path
2
+ dev_0001,qi_linear,dev_c0001,1,500,-0.5260045580824078,dev/tables/dev_0001.csv
3
+ dev_0002,qi_trigonometric,dev_c0002,1,500,2.1639197099173533,dev/tables/dev_0002.csv
4
+ dev_0003,cui_linear,dev_c0003,1,500,1.8174320905943617,dev/tables/dev_0003.csv
5
+ dev_0004,qi_linear,dev_c0004,1,500,-1.402430480710108,dev/tables/dev_0004.csv
6
+ dev_0005,qi_quadratic,dev_c0005,1,500,0.2729301900492508,dev/tables/dev_0005.csv
7
+ dev_0006,cui_linear,dev_c0006,1,500,2.0309497397252354,dev/tables/dev_0006.csv
8
+ dev_0007,cui_linear,dev_c0007,1,500,2.122391805451751,dev/tables/dev_0007.csv
9
+ dev_0008,qi_trigonometric,dev_c0008,1,500,0.1816066216394192,dev/tables/dev_0008.csv
10
+ dev_0009,qi_linear,dev_c0009,1,500,0.9169259184083434,dev/tables/dev_0009.csv
11
+ dev_0010,cui_linear,dev_c0010,1,500,1.374868676884486,dev/tables/dev_0010.csv
12
+ dev_0011,cui_linear,dev_c0011,1,500,1.9949523192290295,dev/tables/dev_0011.csv
13
+ dev_0012,qi_linear,dev_c0012,1,500,-2.3339284159995306,dev/tables/dev_0012.csv
14
+ dev_0013,qi_quadratic,dev_c0013,1,500,-0.07399889623335532,dev/tables/dev_0013.csv
15
+ dev_0014,cui_linear,dev_c0014,1,500,2.000138215614757,dev/tables/dev_0014.csv
16
+ dev_0015,qi_linear,dev_c0015,1,500,-1.8949801249972877,dev/tables/dev_0015.csv
17
+ dev_0016,qi_linear,dev_c0016,1,500,1.2280989356543706,dev/tables/dev_0016.csv
catalog/test.csv ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ table_id,preset,config_id,replicate,n,truth,archive_path
2
+ test_0001,cui_linear,test_c0001,1,500,1.1746063899458445,test/tables/test_0001.csv
3
+ test_0002,qi_trigonometric,test_c0002,1,500,-0.26990319754532444,test/tables/test_0002.csv
4
+ test_0003,cui_linear,test_c0003,1,500,2.364382462638588,test/tables/test_0003.csv
5
+ test_0004,qi_quadratic,test_c0004,1,500,-1.3713871941678224,test/tables/test_0004.csv
6
+ test_0005,qi_linear,test_c0005,1,500,0.9289045325894258,test/tables/test_0005.csv
7
+ test_0006,qi_linear,test_c0006,1,500,1.2062255811707003,test/tables/test_0006.csv
8
+ test_0007,qi_trigonometric,test_c0007,1,500,1.4212794962515538,test/tables/test_0007.csv
9
+ test_0008,qi_linear,test_c0008,1,500,1.08771109916182,test/tables/test_0008.csv
10
+ test_0009,qi_quadratic,test_c0009,1,500,-0.09342755175474472,test/tables/test_0009.csv
11
+ test_0010,cui_linear,test_c0010,1,500,1.555117458792478,test/tables/test_0010.csv
12
+ test_0011,qi_linear,test_c0011,1,500,-0.013740471334333797,test/tables/test_0011.csv
13
+ test_0012,cui_linear,test_c0012,1,500,1.364192579026273,test/tables/test_0012.csv
14
+ test_0013,qi_trigonometric,test_c0013,1,500,0.9186013308829069,test/tables/test_0013.csv
15
+ test_0014,qi_linear,test_c0014,1,500,0.6201577887882804,test/tables/test_0014.csv
16
+ test_0015,qi_quadratic,test_c0015,1,500,0.1828262513573512,test/tables/test_0015.csv
17
+ test_0016,cui_linear,test_c0016,1,500,2.904894264355623,test/tables/test_0016.csv
18
+ test_0017,qi_trigonometric,test_c0017,1,500,0.4064943065456027,test/tables/test_0017.csv
19
+ test_0018,qi_trigonometric,test_c0018,1,500,-0.22402717813993087,test/tables/test_0018.csv
20
+ test_0019,cui_linear,test_c0019,1,500,2.853174147625918,test/tables/test_0019.csv
21
+ test_0020,qi_linear,test_c0020,1,500,2.411874629079611,test/tables/test_0020.csv
22
+ test_0021,qi_quadratic,test_c0021,1,500,0.7203373438954952,test/tables/test_0021.csv
23
+ test_0022,qi_trigonometric,test_c0022,1,500,0.9548288307726619,test/tables/test_0022.csv
24
+ test_0023,qi_quadratic,test_c0023,1,500,-0.3824720330532466,test/tables/test_0023.csv
25
+ test_0024,qi_quadratic,test_c0024,1,500,0.551748745240755,test/tables/test_0024.csv
26
+ test_0025,cui_linear,test_c0025,1,500,2.453629418522129,test/tables/test_0025.csv
27
+ test_0026,cui_linear,test_c0026,1,500,1.08650639718151,test/tables/test_0026.csv
28
+ test_0027,cui_linear,test_c0027,1,500,2.3092470003931953,test/tables/test_0027.csv
29
+ test_0028,qi_linear,test_c0028,1,500,-1.3081337055754256,test/tables/test_0028.csv
30
+ test_0029,qi_trigonometric,test_c0029,1,500,-1.0827935373953468,test/tables/test_0029.csv
31
+ test_0030,qi_trigonometric,test_c0030,1,500,0.09749588133205807,test/tables/test_0030.csv
32
+ test_0031,qi_trigonometric,test_c0031,1,500,-1.5775450012348973,test/tables/test_0031.csv
33
+ test_0032,qi_trigonometric,test_c0032,1,500,-1.3987128492408973,test/tables/test_0032.csv
catalog/train.csv ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ table_id,preset,config_id,replicate,n,truth,archive_path
2
+ train_0001,qi_linear,train_c0001,1,500,-2.8067542739908466,train/tables/train_0001.csv
3
+ train_0002,cui_linear,train_c0002,1,500,2.2001425909188166,train/tables/train_0002.csv
4
+ train_0003,qi_trigonometric,train_c0003,1,500,0.7015672128093834,train/tables/train_0003.csv
5
+ train_0004,qi_trigonometric,train_c0004,1,500,0.16569064602167857,train/tables/train_0004.csv
6
+ train_0005,qi_linear,train_c0005,1,500,0.6276317696616462,train/tables/train_0005.csv
7
+ train_0006,qi_linear,train_c0006,1,500,0.7153684088776275,train/tables/train_0006.csv
8
+ train_0007,qi_linear,train_c0007,1,500,-0.07753245482979915,train/tables/train_0007.csv
9
+ train_0008,qi_trigonometric,train_c0008,1,500,0.7963879260920479,train/tables/train_0008.csv
10
+ train_0009,cui_linear,train_c0009,1,500,2.715552103639384,train/tables/train_0009.csv
11
+ train_0010,qi_linear,train_c0010,1,500,0.7674317763008519,train/tables/train_0010.csv
12
+ train_0011,qi_trigonometric,train_c0011,1,500,0.25600149154175755,train/tables/train_0011.csv
13
+ train_0012,qi_quadratic,train_c0012,1,500,-1.8236759788981616,train/tables/train_0012.csv
14
+ train_0013,cui_linear,train_c0013,1,500,2.2486027716212362,train/tables/train_0013.csv
15
+ train_0014,qi_trigonometric,train_c0014,1,500,-0.7206093422571574,train/tables/train_0014.csv
16
+ train_0015,qi_linear,train_c0015,1,500,-0.9374015072731037,train/tables/train_0015.csv
17
+ train_0016,qi_linear,train_c0016,1,500,-1.5497113282655317,train/tables/train_0016.csv
18
+ train_0017,qi_trigonometric,train_c0017,1,500,0.2534757989390442,train/tables/train_0017.csv
19
+ train_0018,qi_quadratic,train_c0018,1,500,-0.24702175391317321,train/tables/train_0018.csv
20
+ train_0019,qi_linear,train_c0019,1,500,0.7471679338473245,train/tables/train_0019.csv
21
+ train_0020,qi_quadratic,train_c0020,1,500,0.35173816400214014,train/tables/train_0020.csv
22
+ train_0021,qi_linear,train_c0021,1,500,-0.6690686842181339,train/tables/train_0021.csv
23
+ train_0022,qi_trigonometric,train_c0022,1,500,-0.33396944294459907,train/tables/train_0022.csv
24
+ train_0023,cui_linear,train_c0023,1,500,1.730270659659705,train/tables/train_0023.csv
25
+ train_0024,qi_trigonometric,train_c0024,1,500,0.36409426827774455,train/tables/train_0024.csv
26
+ train_0025,qi_linear,train_c0025,1,500,0.4525372993559066,train/tables/train_0025.csv
27
+ train_0026,qi_trigonometric,train_c0026,1,500,0.14250976664013942,train/tables/train_0026.csv
28
+ train_0027,qi_trigonometric,train_c0027,1,500,-0.981040588364424,train/tables/train_0027.csv
29
+ train_0028,qi_linear,train_c0028,1,500,0.8951554385365489,train/tables/train_0028.csv
30
+ train_0029,qi_linear,train_c0029,1,500,4.266913663691322,train/tables/train_0029.csv
31
+ train_0030,qi_trigonometric,train_c0030,1,500,1.546456115731816,train/tables/train_0030.csv
32
+ train_0031,cui_linear,train_c0031,1,500,1.3292556742084942,train/tables/train_0031.csv
33
+ train_0032,qi_linear,train_c0032,1,500,0.6369054490538225,train/tables/train_0032.csv
34
+ train_0033,qi_trigonometric,train_c0033,1,500,-1.6131916646640083,train/tables/train_0033.csv
35
+ train_0034,qi_trigonometric,train_c0034,1,500,0.6584424156331543,train/tables/train_0034.csv
36
+ train_0035,cui_linear,train_c0035,1,500,1.9925965032711008,train/tables/train_0035.csv
37
+ train_0036,qi_quadratic,train_c0036,1,500,-1.2726532394715657,train/tables/train_0036.csv
38
+ train_0037,qi_linear,train_c0037,1,500,0.3792900655177172,train/tables/train_0037.csv
39
+ train_0038,cui_linear,train_c0038,1,500,1.7129762519448488,train/tables/train_0038.csv
40
+ train_0039,qi_linear,train_c0039,1,500,1.0370569008362618,train/tables/train_0039.csv
41
+ train_0040,qi_quadratic,train_c0040,1,500,0.38117773869308236,train/tables/train_0040.csv
42
+ train_0041,cui_linear,train_c0041,1,500,2.2433066802513615,train/tables/train_0041.csv
43
+ train_0042,qi_linear,train_c0042,1,500,-3.5812067721781617,train/tables/train_0042.csv
44
+ train_0043,qi_quadratic,train_c0043,1,500,-0.03877967963795098,train/tables/train_0043.csv
45
+ train_0044,qi_quadratic,train_c0044,1,500,-0.3287975963647826,train/tables/train_0044.csv
46
+ train_0045,cui_linear,train_c0045,1,500,2.8531953431075783,train/tables/train_0045.csv
47
+ train_0046,qi_quadratic,train_c0046,1,500,0.13089830495106275,train/tables/train_0046.csv
48
+ train_0047,cui_linear,train_c0047,1,500,1.966592048631387,train/tables/train_0047.csv
49
+ train_0048,qi_trigonometric,train_c0048,1,500,1.5043882715957144,train/tables/train_0048.csv
official_schema.csv ADDED
The diff for this file is too large to render. See raw diff
 
proximal-official-inputs-v1.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:01603b2c869715ec9486b9a0e73696b632d009a86c77a0f02d1dc69e387bafd6
3
+ size 50129186
proximal-starter-v1.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:55bc8e29b91bdcc440a3888a28db57644cd415558836a74d5096c2b4956bb079
3
+ size 2376155
verify_release.py ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Verify public archives and official-input structure; no references/models needed."""
2
+ from pathlib import Path
3
+ import csv,hashlib,io,json,math,zipfile
4
+ ROOT=Path(__file__).resolve().parent
5
+ def read_csv(data):return list(csv.DictReader(io.StringIO(data.decode('utf-8'))))
6
+ def main():
7
+ checked=0
8
+ for line in (ROOT/'SHA256SUMS').read_text().splitlines():
9
+ expected,relative=line.split(' ',1);p=(ROOT/relative).resolve()
10
+ assert p.is_relative_to(ROOT.resolve()) and hashlib.sha256(p.read_bytes()).hexdigest()==expected,relative
11
+ checked+=1
12
+ schema=read_csv((ROOT/'official_schema.csv').read_bytes());ids=[x['table_id'] for x in schema]
13
+ assert len(ids)==len(set(ids))==2000
14
+ with zipfile.ZipFile(ROOT/'proximal-official-inputs-v1.zip') as z:
15
+ assert z.testzip() is None
16
+ allowed={'README.md','schema.csv','queries.csv','predictions_template.csv','load_observed.py'}
17
+ allowed|={folder+'/'+tid+'.csv' for tid in ids for folder in ['tables','folds']}
18
+ assert set(z.namelist())==allowed
19
+ assert z.read('schema.csv')==(ROOT/'official_schema.csv').read_bytes()
20
+ query=read_csv(z.read('queries.csv'));pred=read_csv(z.read('predictions_template.csv'))
21
+ assert [x['table_id'] for x in query]==ids==[x['table_id'] for x in pred]
22
+ assert all(set(r)=={'table_id','estimate','ci_lower','ci_upper'} and all(r[k]=='' for k in ['estimate','ci_lower','ci_upper']) for r in pred)
23
+ for tid in ids:
24
+ rows=read_csv(z.read('tables/'+tid+'.csv'));folds=read_csv(z.read('folds/'+tid+'.csv'))
25
+ assert len(rows)==len(folds)==500 and set(rows[0])=={'X1','X2','Z','W','A','Y'}
26
+ assert all(all(math.isfinite(float(v)) for v in r.values()) for r in rows)
27
+ assert set(r['A'] for r in rows)=={'0','1'}
28
+ labels=[int(r['fold']) for r in folds];assert set(labels)==set(range(5))
29
+ for f in range(5):
30
+ assert labels.count(f)==100
31
+ assert {r['A'] for r,l in zip(rows,labels) if l==f}=={'0','1'}
32
+ with zipfile.ZipFile(ROOT/'proximal-starter-v1.zip') as z:
33
+ assert z.testzip() is None
34
+ all_ids=set()
35
+ for split,n in [('train',48),('dev',16),('test',32)]:
36
+ truth=read_csv(z.read(split+'/truth.csv'));queries=read_csv(z.read(split+'/queries.csv'))
37
+ catalog=read_csv((ROOT/'catalog'/f'{split}.csv').read_bytes())
38
+ assert len(truth)==len(queries)==len(catalog)==n
39
+ assert {r['table_id'] for r in truth}=={r['table_id'] for r in queries}=={r['table_id'] for r in catalog}
40
+ truth_by_id={r['table_id']:float(r['truth']) for r in truth}
41
+ assert all(math.isfinite(float(r[k])) for r in queries for k in ['X1','X2'])
42
+ for r in catalog:
43
+ assert r['table_id'] not in all_ids;all_ids.add(r['table_id'])
44
+ assert math.isfinite(float(r['truth']))
45
+ assert float(r['truth'])==truth_by_id[r['table_id']]
46
+ observed=read_csv(z.read(r['archive_path']))
47
+ assert len(observed)==500 and set(observed[0])=={'X1','X2','Z','W','A','Y'}
48
+ assert all(math.isfinite(float(v)) for rr in observed for v in rr.values())
49
+ assert set(rr['A'] for rr in observed)=={'0','1'}
50
+ assert not all_ids.intersection(ids)
51
+ print(json.dumps(dict(passed=True,checksums_verified=checked,official_tables=2000,n=500,official_fold_rows=1000000,
52
+ public_starter_tables=96,official_truth_files_present=False,model_or_DGP_calls=0)))
53
+ if __name__=='__main__':main()