junior-ml-dev

Home / example runs

German credit (OpenML 31)

1000 loan applicants, 20 mixed numeric and categorical columns, predict bad credit.

Pre-computed on 2026-11-09 18:19 UTC on a developer laptop (Windows 11), CPU only. Scripted planner, no LLM, no network calls during the run. Dataset: OpenML 31.
Task
binary, target bad_credit, metric roc_auc, 1000 rows x 21 columns
Status
done
Chosen model
linear
Validation roc_auc
0.7913
Reviewer verdict
APPROVE
Columns dropped for leakage
none
Steps
13
Wall time
46.0 s (process start to report); agent clock 43.0 s. One run on a shared machine that was busy with other jobs: a timing sample, not a benchmark.
Budgets
soft wall 55 s (agent stops searching), hard limit 90 s (process killed), FLAML 15 s, 3 tuning trials

Agent steps

#toolserveroklatency (s)
0profile_datasetdata-toolsyes0.09
1detect_leakagedata-toolsyes18.05
2decide_leakageagentyes0.00
3train_modelexperimentsyes0.17
4train_modelexperimentsyes0.48
5train_modelexperimentsyes0.92
6train_modelexperimentsyes3.73
7tune_modelexperimentsyes2.02
8train_modelexperimentsyes16.02
9compare_modelsexperimentsyes0.05
10reviewagentyes0.00
11explain_modelexperimentsyes1.24
12reportagentyes0.00

The agent's own report

Open the report on its own page.