Ambig-DS-T: Target Ambiguity Benchmark

A benchmark for measuring how well data-science agents handle ambiguous prediction targets in tabular Kaggle competitions.

Each task is a Kaggle competition derived from DSBench. For every task we provide two prompt variants — one in which the target column is named, and one in which the target is hidden behind two candidate columns. The agent must select and predict the true target; submissions are graded by the original competition metric using DSBench's per-task evaluator.

The benchmark contains 51 paired tasks (33 classification, 18 regression).

Variants

Variant	File	Description
Full	tasks/{slug}/task.txt	Original task description — names the target column verbatim and lists features under their semantic names.
Ambiguous	tasks/{slug}/task_ambig.txt	Same description with target identity hidden: feature names are anonymized to f_01, f_02, …, the original target name is removed, and training data exposes two candidate target columns val_1 and val_2. Exactly one is the true target; the other is a feature-predictable decoy with the same marginal distribution.

The Full arm establishes the upper baseline; the Ambiguous arm is the diagnostic condition. The failure mode the benchmark measures: an agent that silently picks the decoy column will train a model with normal-looking cross-validation behaviour but score poorly on the held-out test, because the decoy is constructed to be approximately uncorrelated with the true target.

Layout

tasks/
  {slug}/
    task.txt              # Full task description (target named)
    task_ambig.txt        # Target-redacted version (val_1 + val_2 decoy)
    eval.py               # Per-task evaluator (DSBench CLI; metric varies by task)
    _manifest.json        # Provenance + decoy recipe + diagnostics
tasks.csv                 # Flat 51-row index (slug, task_type, n_train, n_test, oracle target column, …)

Setup: getting the competition data

This dataset contains prompts, evaluators, and decoy recipes only — not the Kaggle competition data (train/test CSVs). To respect each competition's terms of use, the redistributed contents are deterministic recipes over data the user must download themselves.

To run the benchmark:

1. Accept each competition's rules in your browser (the source.rules_url field of _manifest.json links straight there).
2. Download the data via the official Kaggle CLI (kaggle competitions download -c <slug>).
3. Use the build script (published separately on GitHub) to apply the deterministic recipe in _manifest.json.ambig_recipe. This rebuilds the ambig CSVs bit-identically using the recorded seeds.master and seeds.decoy.

Manifest

tasks/{slug}/_manifest.json records, per task:

Section	Purpose
source	platform, url, rules_url, wave (all 51 are dsbench_original).
task	task_type (classification/regression), id_column, true_target_column_in_ambig (oracle: val_1 or val_2), decoy_column_in_ambig, original_target_name, n_features, n_train, n_test.
ambig_recipe	Deterministic decoy generation. method (e.g. rank_map_lowcorr_pool+label_noise), full feature_map (original→anon), anon_feature_columns, decoy_pool_anon_features and their decoy_pool_abs_spearman_with_truth, label-noise fractions, seeds.master / seeds.decoy.
diagnostics	cv_true, cv_decoy, cv_ratio_decoy_over_true, correlation_truth_vs_decoy, marginal_match_exact. The decoy is calibrated so that its values are approximately uncorrelated with the true target while remaining roughly as feature-predictable.
eval	Pointer to the local eval.py and its CLI signature.

The task.true_target_column_in_ambig field is the clarification oracle's source of truth: in the clarify experimental condition, an answerer LLM resolves the agent's questions about which column to predict using only this field. It is intentionally never given to the agent in the ambig (no-clarify) condition.

Example task and diagnostics (from bike-sharing-demand):

{
  "task": {
    "task_type": "regression",
    "id_column": "datetime",
    "true_target_column_in_ambig": "val_1",
    "decoy_column_in_ambig": "val_2",
    "original_target_name": "count",
    "n_features": 8,
    "n_train": 8708,
    "n_test": 2178
  },
  "diagnostics": {
    "cv_true": 0.34,
    "cv_decoy": 0.64,
    "cv_ratio_decoy_over_true": 1.90,
    "correlation_truth_vs_decoy": 0.01,
    "marginal_match_exact": true
  }
}

cv_decoy matches cv_true in feature-predictability (median ratio across the benchmark $\approx 1.0$) while correlation_truth_vs_decoy ≈ 0 (median $|\rho_{\mathrm{Spearman}}| = 0.017$): the two columns are equally learnable but almost orthogonal, so an agent that picks the wrong column still gets a high CV score on data that is unrelated to the true target.

Evaluating a submission

Every eval.py accepts the same DSBench-style CLI:

python eval.py --answer_file data/test_answer.csv \
               --predict_file my_submission.csv \
               --path out --name <slug>

…and writes a single float (the competition's original metric — RMSLE / AUC / RMSE / accuracy / …) to out/<slug>/result.txt.

Tasks (51)

#	Competition	Type	True column	n_train	n_test	n_features
1	bike-sharing-demand	regression	val_1	8,708	2,178	8
2	cat-in-the-dat	classification	val_1	240,000	60,000	23
3	cat-in-the-dat-ii	classification	val_2	480,000	120,000	23
4	dont-overfit-ii	classification	val_1	200	50	300
5	instant-gratification	classification	val_2	209,715	52,429	256
6	playground-series-s3e1	regression	val_1	29,709	7,428	8
7	playground-series-s3e10	classification	val_1	94,051	23,513	8
8	playground-series-s3e11	regression	val_1	288,268	72,068	15
9	playground-series-s3e12	classification	val_1	331	83	6
10	playground-series-s3e13	classification	val_2	565	142	64
11	playground-series-s3e14	regression	val_2	12,231	3,058	16
12	playground-series-s3e16	regression	val_2	59,240	14,811	8
13	playground-series-s3e17	classification	val_2	109,143	27,286	12
14	playground-series-s3e2	classification	val_2	12,243	3,061	10
15	playground-series-s3e20	regression	val_1	63,218	15,805	75
16	playground-series-s3e22	classification	val_1	988	247	27
17	playground-series-s3e23	classification	val_1	81,410	20,353	21
18	playground-series-s3e24	classification	val_1	127,404	31,852	22
19	playground-series-s3e25	regression	val_1	8,325	2,082	11
20	playground-series-s3e3	classification	val_2	1,341	336	33
21	playground-series-s3e4	classification	val_2	175,303	43,826	30
22	playground-series-s3e5	classification	val_1	1,644	412	11
23	playground-series-s3e6	regression	val_2	18,184	4,546	16
24	playground-series-s3e7	classification	val_2	33,680	8,420	17
25	playground-series-s3e8	regression	val_1	154,858	38,715	9
26	playground-series-s3e9	regression	val_1	4,325	1,082	8
27	playground-series-s4e1	classification	val_1	132,027	33,007	12
28	playground-series-s4e2	classification	val_2	16,606	4,152	16
29	playground-series-s4e4	classification	val_2	72,492	18,123	8
30	playground-series-s4e5	regression	val_2	500,000	223,592	20
31	playground-series-s4e6	classification	val_1	61,214	15,304	36
32	porto-seguro-safe-driver-prediction	classification	val_2	476,169	119,043	57
33	santander-customer-satisfaction	classification	val_1	60,816	15,204	369
34	santander-customer-transaction-prediction	classification	val_1	160,000	40,000	201
35	santander-value-prediction-challenge	regression	val_1	3,567	892	4,991
36	spaceship-titanic	classification	val_1	6,954	1,739	12
37	tabular-playground-series-apr-2021	classification	val_2	80,000	20,000	10
38	tabular-playground-series-aug-2021	regression	val_1	200,000	50,000	100
39	tabular-playground-series-aug-2022	classification	val_1	21,256	5,314	24
40	tabular-playground-series-feb-2021	regression	val_1	240,000	60,000	24
41	tabular-playground-series-feb-2022	classification	val_2	66,900	16,725	286
42	tabular-playground-series-jan-2021	regression	val_1	240,000	60,000	14
43	tabular-playground-series-mar-2021	classification	val_1	240,000	60,000	30
44	tabular-playground-series-mar-2022	regression	val_1	500,000	169,767	4
45	tabular-playground-series-may-2022	classification	val_1	720,000	180,000	31
46	tabular-playground-series-nov-2021	classification	val_2	480,000	120,000	100
47	tabular-playground-series-oct-2021	classification	val_2	500,000	200,000	285
48	tabular-playground-series-sep-2021	classification	val_2	416,946	104,237	118
49	titanic	classification	val_1	712	179	10
50	tmdb-box-office-prediction	regression	val_1	2,400	600	21
51	ventilator-pressure-prediction	regression	val_1	500,000	1,207,200	6

The True column column is the answer the clarification oracle returns when an agent asks which of val_1/val_2 to predict; it is never shown to the agent in the ambig condition.

Citation

@article{ambig-ds-2026,
  title  = {Ambig-DS: Diagnosing Unflagged Misframings in Data-Science Agents},
  year   = {2026},
  note   = {NeurIPS 2026 Datasets \& Benchmarks submission (under review)}
}

License

The contents of this repository (prompts, manifests, task index, decoy recipes) are released under CC-BY-NC-4.0, inheriting the non-commercial research-use restriction from the upstream DSBench dataset terms (DSBench code is MIT). The task.txt files are factual paraphrases of publicly available Kaggle competition descriptions; the task_ambig.txt files, the decoy-generation recipes in _manifest.json.ambig_recipe, and the per-task diagnostics are original contributions.

The per-task eval.py evaluators are redistributed unchanged from DSBench (Jing et al., 2024) so that grading remains bit-identical to upstream. Some still contain inline comments in Chinese — these are upstream artefacts.

The underlying Kaggle competition datasets are not redistributed here. They must be downloaded separately via the Kaggle API and remain subject to each competition's individual rules and terms of use.