remove not needed path

5g_energy_consumption/019db4c5-d45f-7c07-80c9-bdd75ea01338/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\n- Note: The dataset was used in a Zindi challenge, but also uploaded to Huggingface under a MIT license by the company (Huawei).\n- The corresponding ITU/Zindi challenge explicitly emphasizes generalization to unseen base station products/configurations. We therefore split by base_station (BS).\n- With this setup, the task is development of predictive models for network optimization where models need to generalize to new unseen base station configurations and estimate their energy consumption under similar conditions (during the same time period).\n- For preprocessing, we orient on a top solution from the Zindi challenge: https://github.com/ITU-AI-ML-in-5G-Challenge/5G-Energy-Consumption-Modelling-Solution-Team-Farzi-Data-Scientists/tree/main.\n- Note that the competition also used mostly samples from the known BS as the test set, but weighted unknown base stations higher in evaluation.\n- The data itself is time-series. However, because we predict entirely unseen base stations in a per-sample fashion, the row-wise dependencies resulting from the temporal components cannot be used to improve performance unless the task is treated as transductive learning.\n- Following the Zindi solution, we merge the three given tables and use the Cell0 information only from the cell level table. In addition, we merge the Cell1 information since it contains information as well which might be useful in a grouped split setting.\n- Following the Zindi solution, we derive calendar features (`day`, `hour`, `weekday`) and drop the absolute timestamp.\n- We drop constant columns.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

acquire_valued_shoppers_challenge/019d7376-a3c5-7df7-88c0-ba4e87c3dfa0/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe follow the preprocessing from TabRed https://github.com/yandex-research/tabred/tree/main/preprocessing#ecom-offers-acquire-valued-shoppers-by-dmdave (which follows a top solution https://github.com/MLWave/kaggle_acquire-valued-shoppers-challenge).\n\n- The preprocessing by TabRed collapses customer groups into single entries per customer. So we go from grouped-temporal data to only temporal data. Future work could look into a version without preprocessing.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

airfoil_self_noise/019d7366-4673-7d9a-9821-35604af616f6/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "N/A",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

allstate_claims_severity/019d736a-2321-76da-a36a-ddc5e1fbf7b6/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\nWe start with the train.csv from Kaggle.\n\n- The data has been anonymized.\n- Top Kaggle solutions did not perform any relevant preprocessing.\n- We drop the ID column, as it does not contain any signal.\n- The data contains 1 duplicate row when ignoring the target. We drop this artifact.\n- We log scale the target.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

amazon_employee_access/019d7366-58b8-7e97-bf34-83c96a915561/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We only use the training data from Kaggle.\n- We renamed the target \"ACTION\" to \"ResourceApproved\" and mapped binary values to \"Yes\"/\"No\".\n- Anomaly: the data might contain sub-groups related to managers and resources.\n- Anomaly: likely, similar to the test data, each sample represents a unique employee.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

amex_non_iid/versions/019d7455-0e4e-7261-9842-93177684d486/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe start with the raw data from Kaggle and do not apply any further preprocessing to simulate a pipeline that can handle raw non-IID grouped data.\n",
   "version_from_unique_name": "amex_non_iid",
   "version_comment": "\nWe randomly sub-sample to 1.25 million rows such that train and test splits are roughly 1M and 250k rows respectively.\n",
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

anes_voting_2026/019db4ba-6124-7a56-a0b5-9df8811adc2f/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\n- We remove all samples with missing pre-election data, missing post-election data or a missing target.\n- We drop 327 that are missing from the last 9 years, which will be used for the test splits.\n- We drop all columns with questions that appear only in one year.\n- The question columns partially stem from post-election interviews. However, there is no simple way to tell which is post-election. We therefore drop all questions for which \"no post IW\" is listed as a reason for missing values. Some additional questions are dropped based on common sense.\n- We encode gender as a categorical variable, because Other was introduced in 2016. However there are very few of these samples. This would actually require split-specific preprocessing.\n- For all features, we assign \" \" as missing values. Most features have additional types of missingness with own codes. We keep them as categories. That means NA only is assigned if a question was missing in a survey.\n- We transform only high-cardinality (>10) numeric features to numeric, and leave all other ordinal features as categorical, since they all include at least one value that is out-of-order (e.g., \"Don't know\", \"Refused\", \"Not applicable\", \"Other\", etc.).\n- Although there are several ways to further preprocess features, we leave the preprocessing minimal to enable the development of models that can handle even complex preprocessing on their own.\n- Note: Alternative target could be constructed using who the candidate voted for (VCF0704). This could be a multi-class target with [Democrat, Republican, Other Party, Did not vote, Voted but unknown who]. However, the last category is less useful to predict, and might introduce problems. We therefore stick to predicting whether someone voted or not.\n- Note: While the data was also used in the TableShift benchmark, we define an entirely different task, with a temporal split and more features.\n- Note: Some information is represented across multiple columns, e.g.,  ['VCF0009x', 'VCF0009y', 'VCF0009z', 'VCF0010x', 'VCF0010y', 'VCF0010z', 'VCF0011x', 'VCF0011y', 'VCF0011z'] all correspond to weight. We keep all of these columns, even though some might be redundant, because we want to keep the data as close to the original as possible leaving the challenge to the model.\n- Note: Ideally, we would select different feature sets per data split, depending on feature availability. However, for simplicity, we just keep all features.\n- Note: We considered keeping the unique respondent identifier in the data, because otherwise models that are able to reconstruct this information would unfairly benefit. However, for now we drop the identifier since it is not a desirable predictive feature for the given task, because the goal is generalization to the population, not memorizing individual behavior.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

aps_failure/019d7366-7373-7942-9438-aa122a6ca492/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\n  - We combined the original training and testing data into a single dataset.\n  - We renamed the target feature to \"AirPressureSystemFailure\".\n  - We converted \"na\" strings to real NaN/missing values, making the data numeric.\n  - Anomaly: we cannot determine the data types of the features.\n  - Anomaly: some features are bins of histograms (see original data description).\n  - Anomaly: the original task used a cost matrix for evaluation.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

asp_potassco_classification/019d73ec-935f-7f88-8a82-af3ad2f44e9f/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\nWe get the data from ASlib and merge them into one file.\n\n- We treat it as a multiclass classification task to solve the algorithm selection task as in the OpenML version (https://openml.org/d/41705).\n- We drop all cases where no algorithm was able to finish before the timeout as these are essentially random labels.\n- We resole the instance ID to a task ID by mapping the instance ID back to the source task based on their naming convention. From our understanding, the data contains multiple instance from the same solver task, differing only in seed or task configurations. We want to avoid having samples from the same task in both train and test set, as this would leak what algorithms are best for this task, and since in real-world you might not have samples from your new task. Thus, we treat the data as a grouped task, where we aim to generalize the predicting the best algorithm to new tasks across a set of instances of this task.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

audiology_diagnosis/019d736a-3b3d-73da-8d65-b4b7a39e1590/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the data from UCI and merge train and test data.\n\n- We drop the ID column as it is uninformative here.\n- The target label contains groups of labels with specifications. We merge labels into groups that represent general a diagnosis. We create 3 labels: \"normal\", \"cochlear\", \"other\". There is likely a much better way to partition these labels, but this is most reasonable partitioning for an acutal task, going from names and my limited domain knowledge about the various diagnoses.\n- We drop duplicated rows as they might introduce too much information leakage for such small data and are likely not natural but rather an artifact of the limited number of features.\n- We remove one constant column (history_fullness).\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

bad_customer_detection/019d7366-8875-7ee1-8f3b-5ad96bdd2324/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We renamed the values of the target variable to be more descriptive.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

bank_customer_churn/019d7366-99cc-77cc-8d41-5c3ccc8eff96/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We remove the customer_id column.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

bank_marketing/019d7366-abe0-7dcc-a9fb-069bfb233c78/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We removed the \"duration\" feature following its original description to obtain a \"realistic predictive model\".\n- We further remove the \"month\" and \"day_of_week\" features, as they also relate to the last contact -- which is not available in a real-world scenario.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

biogeographical_ancestry_prediction/019d7374-b899-714c-9372-37a48f580b0f/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe use this dataset as one of the most recent example of a machine learning task based on the Human Genome project.\nWe take the targets from the paper by Heinzel et al. (2025) and only rename the targets to be standardized and more concise.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

biomechanical_orthopaedic_prediction/019d736a-4c2d-74b6-95a3-c63bd4b8fff7/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the UCI version, which is the oldest existing version of the dataset we found.\n\n-\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

bioresponse/019d7366-c770-7d3c-8dd9-b9cc456542d6/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We changed the name of the target and mapped it to yes/no\n- Anomaly: Only train data is used, since test data is empty for target\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

blood_tests_drink_prediction/019d736a-5d45-743a-b7c5-11da6b494fb7/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nThe task is framed as a liver disorder prediction task, but we only have data bout the amount of drinks consumed. So instead, we will frame it as a task to predict the number of drinks based on the blood work data. This is a proxy for the original task, where later based on the number of drinks the liver disorder was determined.\n\n- We drop the selector column, as we ignore the original train/test split, and will create our own splits.\n- The data contains natural duplicates, so we keep them.\n- We log1p scale the target.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

blood_transfusion/019d7366-d9d0-777d-b0ae-e7a5175f7f09/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We made feature names more descriptive.\n- We renamed the target and mapped binary values to \"Yes\"/\"No\".\n- Anomaly: the data has a lot of duplicates (29%) and several duplicates with different target values.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

body_density_prediction/019d736a-6e2b-77a5-ae25-1222ef4fe4e9/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the data from Kaggle.\n\n- Note, the task is not about bodyfat prediction as this is determined by a deterministic formula. Instead, the task is to estimate the density (which can be used to get the body fat via the deterministic formula). However, density requires a test. So we aim to predict from data the density such that we can skip this test. Which can make a lot of sense, if the test is too expensive as it requires underwater weighing.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

california_house_prices_2020/019d7388-c562-72b3-a533-5f588dc737c9/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe follow some common preprocessing steps from Kaggle and enhance features as much as possible.\n\nWe found out that the data was web scraped from redfin.com and that the houses are ordered by the time they were sold such that a house with a higher ID was sold later. We confirmed this by looking at the history of various houses on redfin.com that we found in the dataset and checking their order in the ID with the time they were sold. Houses with a higher ID were sold later.\n\n- We log scale the target variable.\n- We do not have the exact dates but we know a higher ID means a later sale. We name the ID column accordingly to \"time_index\" and use it as a time feature.\n- The descriptions did contain the last sold price with a standard phrase such as \"This home last sold for $X in January 2020. The Zestimate for this house is $Y The Rent Zestimate for this home is $Z/mo.\". In these cases $X is identical to the target variable. Given such a drastic data leakage, we remove all rows (5413) that contain this phrase in the description to ensure they are not different in other ways too.\n- Note, sometime descriptions are missing and we intend that models/pipelines need to be able to handle this.\n- When investigating the \"Lot\" column we found many cases where the lot size (given in sq ft) is incorrect compared to the official lot size on the internet. The parser seems to have had an issue because the website often incorrectly showed the sq ft but gave the unit as acres. We found cases that were wrong by checking the acres size of the houses and anything with more than 2000 acres was investigated and subsequently corrected.\n- We parsed the bedrooms with a proxy following Kaggle and added an additional column containing only the semi-free text descriptions of the bedrooms.\n- We fix the parser errors for two cases for total interior livable area that were clear outliers based on the real data on redfin.\n- Garage and total space were in most cases the same values parsed from different fields. We keep Garage space and add an 'Extra Space' column that is the difference between total and garage space as an additional feature. This results in some cases with negative extra space that are most likely parser errors. We set them to nan. Note, we observed that that garage space number is in many cases a parsing error and does not reflect the real garage count.\n- The data has a lot of spatial features that are already semi-resolved to distances. This could be further enhanced using the address information. We leave it to the pipeline to handle this, if at all.\n- The dataset contains many text-like many categorical that are also high-cardinality, so perfect string data.\n- We drop all entries from the State of Arizona (AZ) as they are only a small fraction (n=431 after our preprocessing). The dataset is specified to be for California house prices.\n- We treat ZIP codes as strings. But we add a new categorical features that puts zip code in to approximate relevant buckets of regions.\n- We fixed one entry that had a negative value for garage space, likely due to a typo. This error still exists on the website.\n- After all the above preprocessing, there exist 63 houses that appear multiple times in the dataset (same address, zip code, year built). We only keep the newest entry to avoid group-related target leakage.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

cardiotocography/019d7392-12c2-7a87-991c-13a7246555d4/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\nWe use the data from UCI and the 3 class problem of predicting NSP as it is more medical relevant.\n\n- We transform the file names into patient IDs to indicate the sub-group of samples recorded from the same patient.\n- The raw data has dates and the start and end index of the recording. We drop these as they are not relevant for the predictive task and would not be available for test samples.\n- We drop all other columns that represent the label.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

churn/019d7366-eb2d-72f9-9998-72dfc5b6cc79/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\n- The original source is lost, so we use https://github.com/EpistasisLab/pmlb/tree/master/datasets/churn (or https://www.openml.org/d/40701)\n- We dropped the \"phone_number\" feature as it seems to be an index in the original data. \n- We renamed the target variable to \"CustomerChurned\" and mapped binary variables to \"Yes\"/\"No\"\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

cirrhosis_patient_survival_prediction/019d736a-9116-7589-9cad-4cee540f2926/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the UCI data.\n\nWe note that this is a survival prediction task. But the way we split the data and score the task, it is not treated correctly as a survival prediction task. At the same time, it is randomized control data based on the Drug column. We treat it as a regression task to predict the time-to-death for non-censored patients, including the drug column into the modelling.\n\n- We drop ID and Status column as they have no relevance after filtering to only non-censored patients.\n- We log transform N_days and rename it to the target \"log_days_to_death\".\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

climate_model_weather_forecasting/versions/019d7379-7906-707e-95e9-f479afb2e2c1/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\nWe start with data from TabRed, which already comes preprocessed.\n\n- This data is similar to data from https://arxiv.org/abs/2406.19380. This paper also describes the feature in more detail. The column `apply_time_rl` is new and might relate to a rolling lag/window indicator.\n- We drop duplicated columns `cmc_0_1_66_0` , `cmc_0_1_67_0`, `cmc_0_1_68_0`.\n",
   "version_from_unique_name": "climate_model_weather_forecasting",
   "version_comment": "\nWe randomly sub-sample the train to 1 million and test data 250k rows. We follow TabReD and use random sub-sampling. The idea behind this instead of a time-based subsampling is to keep data from various time periods and model the distribution shift across the full time horizon.\n",
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

clock_protein_toxicity/019d7375-2841-7bfb-a6da-5adfdcd43d98/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We remove duplicated columns (same values for all rows).\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

coffee_rating_prediction/019d7388-dffd-7bc4-ab3d-40bf6581786e/dataset_metadata.dataset-mold-v1.json

@@ -17,6 +17,5 @@
   "curation_comments": "\nWe start with the Kaggle version. Note, we acknowledge that the dataset is not of the highest quality and likely does not represent a real-world prediction task. Nevertheless, it can be used to test models.\n\n- We drop slug as it is a unique identifier that does not have predictive power.\n- We drop desc_2 as it in general includes notes that may or may not be related to the coffee quality. Moreover, it sometimes leaks the rating (e.g. through award notes). We only keep desc_1 and desc_3, which are mostly descriptions of the coffee.\n- We drop all_text as it is a concatenation of desc_1, desc_2, and desc_3. And desc_2 includes data leakage (such as if the coffee won an award or not). Thus, we drop all_test.\n- We drop aroma, acid, body, flavor, aftertaste, and with_milk as they are all features of the rating and thus leaking the target. Any of them could be used as a target, we focus on the overall rating.\n- We parse agtron into a lower and upper value as float.\n- Note, the data contains several spatial features about locations. We do not resolve these.\n- The price feature is extremely messy as a result of web scrapping. We resolve this feature as real tabular data (not from a web scrapper but from the database source), would have these information stored in a more structured way. We drop coffees that are measured in \"capsules\" units as they exist only 8 times and seem to be strong outliers. We normalizes the price to be per grams. Moreover, we convert all currencies to USD (with a fixed rate); we ignore the temporal incorrectness this may introduce as it is most likely minimal.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

coil_2000/019d7366-fdc2-7225-a22c-314c624711f6/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We created semantic meaningful names for the features.\n- We combined the original training and validation data into one new dataset.\n- We reversed the ordinal encoding of the original data where possible.\n- Anomaly: the data has 15% duplicates.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

concrete_compressive_strength/019d7367-0f12-7c25-b411-de9c711bd4ec/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We rename features to be shorter while similar to the original names.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

consumer_complaints/versions/019d738b-4c6e-751f-972a-5f63b1508f70/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nThe dataset on Kaggle (https://www.kaggle.com/datasets/selener/consumer-complaint-database) has data from up until 2019. We use the newest version from the government website with data up until 2025.\n\n- Context and descriptions of the features can be found here: https://cfpb.github.io/api/ccdb/fields.html\n- \"Company public response\" is not free text but selected \"from a set list of options\".\n- Following TexTabBench, we focus on predicting the type of closure a complaint got, that is the \"Company response to consumer\" column. We want to predict if a complaint will be closed with just an explanation, with non-monetary relief, or with monetary relief.\n- We filter the data to only include entries after consumer disputations were discontinued as this represent a shift in protocol. This filters all data before April 24th 2017.\n- We filter all cases where the \"Consumer consent provided?\" is in progress or got an untimely response.\n- We filter all rows that do not include consent to share their complaint narrative. This ensures the data contains text sentences.\n- We drop \"Company public response\" as it leaks the target variable.\n- We only allow complaints from US states (no territories or international complaints) and remove cases with missing states.\n- The data has unresolved spatial information in the ZIP code. Some ZIPs are censored (ending in \"XXX\" or full removed \"XXXXXX\").\n- We add a feature for \"Low population area\", which determines that the ZIP is censored.\n- The tags filed contains only three non-nan labels, of which one is a duplicate of the others. We created two categorical features from it instead.\n- We drop duplicates (2% of the data) as the data should not contain naturally occurring duplicates and this likely results from some overlap in data collection or data entry or faulty re-submissions. We investigated some of the duplicates and they appear to be identical complaints. There exist duplicates with different target labels, we also drop these as we have no way to determine what the correct label is.\n- We drop constant columns, the complaint ID, and when the complaint was send to the company (as it does not related to the target task)\n",
   "version_from_unique_name": "consumer_complaints",
   "version_comment": "\nWe use the last 3 months as test data randomly sub-sample the train to 1 million and test data 250k rows. We follow TabReD and use random sub-sampling. The idea behind this instead of a time-based subsampling is to keep data from various time periods and model the distribution shift across the full time horizon.\n",
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

cooking_time/versions/019d737d-b4b1-7c24-8930-70a2c242b3f9/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe start with data from TabRed, which already comes preprocessed.\n",
   "version_from_unique_name": "cooking_time",
   "version_comment": "\nWe randomly sub-sample the train to 1 million and test data 250k rows. We follow TabReD and use random sub-sampling. The idea behind this instead of a time-based subsampling is to keep data from various time periods and model the distribution shift across the full time horizon.\n",
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

covertype/019d7391-8e33-757b-b38a-79ec188fa001/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe start with the data from UCI.\n\nWe create a special version of the dataset to avoid leakage. This version consists only of three 3 classes and 3 (spatial) wilderness areas. The investigation of Covertype shows that the dataset comprises multiple wilderness areas (Rawah, Comanche Peak, Neota, Cache la Poudre) that can be treated as subgroups but are not strictly IID, with wilderness area and soil type encoded as one-hot categorical features. Although TabRed argues for a time split to reflect a real task, the dataset lacks both explicit time and precise spatial (e.g., GNSS) features, leaving only area identifiers, which implicitly encode collection time and location; consequently, IID splits would introduce temporal or spatial leakage. Additional issues include transformed features in the OpenML/TALENT version that leak test distribution, incorrect column order in the UCI release, and evidence from EDA that class distributions differ across areas, confirming the dataset\u2019s grouped nature where cover type characteristics strongly depend on area. The only robust strategy is a spatially motivated grouped split by area; however, because some classes appear only in specific areas, the dataset is restricted to classes present in the three largest areas and evaluated using leave-one-area-out grouped splits.\n\nOther steps:\n- We add the column names in the correct way.\n- We reverse the one-hot encoding of the Wilderness_Area and Soil_Type features, and add human-readable descriptions for these features. Moreover, we add the climatic and geologic zones categorical features for each soil type based on the codes.\n- We reverse the ordinal encoding of the class name.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

credit_approval/019d736a-a27c-7c85-90c5-ca1d5dc849bd/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe get the data from UCI.\n\n- We encode A14 as numeric, following the description of it being a continuous feature.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

credit_card_clients_default/019d7367-23b0-7125-9a7b-9ad1a90dc75d/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n  - We rename the target variable and restore the original class names.\n  - We drop the \"ID\" column.\n  - Anomaly: the data has temporal features but the task is time-invariant.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

credit_g/019d7367-39a7-7335-8448-2dc9b2872361/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We reversed the original ordinal encoding.\n- Anomaly: the original task used a cost matrix for evaluation.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

customer_satisfaction_in_airline/019d7367-4d56-7274-a470-1609f3a3d6fe/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We renamed the target column from \"satisfaction\" to \"satisfied\" for clarity and mapped the values to \"Yes\" and \"No\"\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

delivery_eta/versions/019d7382-df54-7dd3-8198-0567d7499858/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe start with data from TabRed, which already comes preprocessed.\n\n- We drop a duplicated column `cat_2` in the preprocessing.\n",
   "version_from_unique_name": "delivery_eta",
   "version_comment": "\nWe randomly sub-sample the train to 1 million and test data 250k rows. We follow TabReD and use random sub-sampling. The idea behind this instead of a time-based subsampling is to keep data from various time periods and model the distribution shift across the full time horizon.\n",
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

dementia_prediction/019d7391-61e5-7cca-9a66-e85718233be9/dataset_metadata.dataset-mold-v1.json

@@ -16,6 +16,5 @@
   "curation_comments": "\nWe start with the data from Mendeley.\n\n- CDR and Group are both the same target variable. CDR is the Clinical Dementia Rating (0 = no dementia, 0.5 = very mild AD, 1 = mild AD, 2 = moderate AD), which determines the group variable. The cases for converted / changed their dementia rating over time. We make this a task to predict for a patient (at any given time point of a scan) their dementia rating. The rating is ordinal but discrete, so we treat it as a classification problem (not regression). We drop cases with moderate_AD (n=3), as we do not have enough data on this class to include them in our prediction task.\n- We drop the group variable and any information about the time of the scan as our goal to predict the rating from eTIV, nWBV, and ASF which are all derived from the MRI scan and independent of time.\n- We also drop the constant hand column.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

diabetes_130_us/019d7367-64e4-7883-956d-9ba2fc8db41b/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We drop duplicated patients based on the \"patient_nbr\" feature to avoid target leakage.\n- We reversed the original ordinal encoding for three ID-based features (admission_type_id, discharge_disposition_id, admission_source_id).\n- We created the target from the \"readmitted\" column following the original task description: \"<30\" becomes \"Yes\", everything else \"No\".\n- We dropped \"encounter_id\" and \"patient_nbr\", which are both unique identifiers for each row.\n- We keep original \"?\", NULL-codes, and NaN values because they exist in different ways across the columns.\n- Anomaly: There is a distribution shift based on the original order. The reason for this might be that the encounters are ordered in some way such that later parts of the data contain different sub-groups than earlier parts. This is also indicated by the fact that the \"payer_code\" feature is responsible for the shift. This distribution shift vanishes after randomly shuffling the data (as done by default for this and all other datasets used in TabArena).\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

diamonds/019d7367-779c-707e-b777-849180166b74/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- Unlike TabArena, we log scale the target as it is a price and hence log scaling is generally recommended.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

drug_induced_autoimmunity_prediction/019d736a-b4b1-7f51-9563-dde2b6ecd948/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the dataset from UCI.\n\n- We drop the constant columns.\n- We keep the SMILES code as string for later pipelines to handle.\n\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

early_learning_predictors/019db4ff-b5e9-7b27-80b1-8cf2ca025d20/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\n- The data is collected for predictive analytics to find major drivers of early learning outcomes. Nevertheless, predictive performance matters and the data could also be used for identifying at risk children.\n- The data consists of different subsets that were collected at different times. Sometimes the target was even collected before the predictors. Although the data is collected over time, the task is cross-sectional. Therefore, we drop most date features, after checking whether relevant information was already extracted.\n- We use a grouped split based on id_facility to reflect a potential deployment scenario where we want to predict for a new facility. Note that a date (year) is available, but the data collection is conducted per facility so we don't use a time-based split.\n- The target score ELOM is continuous, but decisions are made based on cutoffs to identify at-risk children. We keep the task as regression, to maintain the most information.\n- Note: The data consists of different age groups, for which the ELOM score is computed differently. While under these circumstances a multi-class target using the given cutoffs might be a valid choice, the age is given  as a feature and therefore, we assume that a model can be expected to learn patterns for subgroups.\n- The data is collected from several sources, including direct assessments, teacher questionnaires, parent surveys and facility data. The data is collected by different enumerators across different facilities. The target is from a separate tool (see further information below).\n- We drop all columns which are components or alternative versions of the target, to avoid leakage.\n- We drop all samples with a missing target.\n- Due to most samples having missing values, we drop all \"HOME LEARNING ENVIRONMENT DATA\" features. This mostly aligns with the Zindi challenge, except that the features hle_ind, quintile_used, and ses_proxy were kept. We keep the latter two, because they are not missing for all samples and are derived from other; not-HLE features.\n- We drop all features with 100% missing values and one unique value.\n- We drop all indicator columns.\n- We drop some repeated columns, such as age_group_label.\n- Because all features could generally be available at prediction time, even if would not be ideal effort-wise, we keep all features that don't leak the target or are otherwise problematic.\n- We transform all date columns to datetime.\n- Additional description of the target: The ELOM 4 & 5 Years Assessment Tool measures whether preschool children are on track for their age in key areas of development. It is a standardised tool that measures performance across five key developmental domains for children aged 50 to 59 months and 60 to 69 months. The scoring assesses a child in 23 items, across five domains, namely: gross motor development, fine motor development, emergent numeracy and mathematics, cognition and executive functioning, and emergent literacy and language. These five domains form part of the direct assessment.\n- Note: We keep some features that were not in the zindi challenge, such as the teacher and parent questionnaires, because they are potentially predictive and could be available at prediction time. Moreover, some columns present in the zindi challenge (like latitude and longitude) are not available.\n- Note: The data consists of a few subsets with partially distinct features collected. This might not represent a real task, but we don't know for sure, so we keep all samples.\n- Note: The enumerator_id is a control variable of the data collection rather than a factor that should be used for prediction. However, it is predictive and explains noise in the target, so we keep it as a feature.\n- Note: There are strong patterns of missingness within the data, due to different uses of various tools and measurements across subgroups\n- Anomaly: Many missing values\n- Anomaly: This is a highly underspecified task, leaving a lot of data cleaning to the user/model.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

early_stage_diabetes_risk_prediction/019d736a-c5ee-7069-9b14-686535f9caa9/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe use the data as is from UCI.\n\n- Note, the dataset contains a lot of naturally occurring duplicates (50%). We drop them to avoid too extreme duplicated-based data leakage biasing the evaluation.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

ecoli_proteins/019d736a-d6d2-7e28-aeff-7da6b7126b9d/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the dataset from UCI.\n\n- We drop the sequence name column, which is a unique identifier.\n- We drop three classes with less than 10 samples each: omL (5), imL (2), and imS (3).\n- We drop \"chg\", which becomes constant after our preprocessing steps.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

ecommerce_shipping/019d7367-8a04-7bf2-8ab2-68d1726cf31e/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We dropped the ID column.\n- We renamed the target feature \"Reached.on.Time_Y.N\" to \"ArrivedLate\" and mapped binary values to \"Yes\"/\"No\".\n- Anomaly: the target and task seems somewhat disconnected from the features. Moreover, some source information on the data is missing and there might be some translation issues.\n- Anomaly: there might be some data issues related to \"Warehouse_block\" and the value \"F\" consisting of two block \"E\" and \"F\".\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

electric_motor_temperature_prediction/019d7392-e21f-7e60-9efb-2141c3513fd9/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\nWe select the task to predict permanent magnet temperature from the input features. We create grouped splits on the profile_id, which corresponds to predicting the temperature for an unseen motor run session, thus simulating how the model would be used in reality.\n\n- The original study also incorporates temporal connection within the session. We re-create a time-index (https://www.kaggle.com/datasets/wkirgsn/electric-motor-temperature/discussion/147446).\n- We also add the derived inputs used in the original study.\n- We also add the EWMA/WEMS features with window sizes of [500, 2000, 4000, 8000]. To avoid the problem of having an incorrect EWMA/WEMS state and since not all recording are warmed up (https://www.kaggle.com/datasets/wkirgsn/electric-motor-temperature/discussion/117319), we drop the first 500 samples (first span) of each profile to simulate such a warm-up.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

emscad/019d736a-ed4a-7aaa-89df-d52558565db9/dataset_metadata.dataset-mold-v1.json

@@ -15,6 +15,5 @@
   "curation_comments": "\nWe have no access to the original state and the data already contains some preprocessed features.\nThe paper describes them in more detail as well in Table 2.\n\nIn the original paper, the authors subsampled the data to have 450 fraudulent and 450 non-fraudulent samples.\nNo details are given on how the subsampling was done. Moreover, duplicates were skipped, but no details are given on how duplicates were identified.\n\n- The dataset was collected from 2012 to 2014. But the data contains no feature to identify the time of a sample. So we can expect that some bias from temporal leakage. Moreover, the data might contain multiple job postings from the same fraudster, which would need to be grouped, but we cannot identify them from the data.\n- Following the original paper, we drop all duplicates (when ignoring the job_id column) to avoid data leakage.\n- The data has a location description, we do not resolve it lat/longitude but leave it to the pipelines.\n- The salary range has several weird quirks. It contains data either in the thousands, or is missing the \"k\" to denote thousands. Moreover, it contains empty or 0-0 entries. The column might contain yearly salary, hourly salary, or one-time payment. Finally, for some jobs, the column contains a date (e.g. \"Oct-20\"). We keep the column as complex as it is. We add one column that contains the maximum salary parsed from the text and we only keep salary values above 50k as valid values to create a numerical feature for differences in the larger ranges.\n- We removed job listings with non english texts (138 rows).\n- We do not subsample the data, as a result, the data is heavily imbalanced.\n- We found no relation of job_id with a timestamp and even found cases where a lower job_id has a description claiming to be a job from 2014 (so the end of the collection period). Thus, we do dropped job_id as it does not contain any signal.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

eryhemato_squamous_disease/019d736b-020a-7fbc-9785-6438802781f2/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\nWe start with the data from UCI.\n\n- We encode all features but age as categorical, since they are ordinal features in nature.\n- We ensure missing values in age are encoded as NaN.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

fiat_500/019d7367-9b82-78a9-a53a-570038a28d0a/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "N/A",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

fitness_club/019d7367-aced-7960-94c9-b19b393cb0ad/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We dropped the booking_id column.\n- We renamed the values of the target variable to be more meaningful (\"Yes\"/\"No\").\n- We treat \"-0\" values as \"0\" values for the target, following the description.\n- We removed trailing words (like \"days\") from \"days_before\".\n- We aligned the naming of \"day_of_week\" to be consistent per day.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}

food_delivery_time/019d7367-beff-7ca7-9033-8f346962f7a1/dataset_metadata.dataset-mold-v1.json

@@ -14,6 +14,5 @@
   "curation_comments": "\n- We dropped entries with a duplicated ID, keeping only the first one.\n- We dropped the ID column.\n- Anomaly: The ID of the delivery person is given as a feature. In some contexts, this feature might not be allowed to use. Moreover, using this information might require considering a temporal split where a cold-start problem is covered.\n",
   "version_from_unique_name": null,
   "version_comment": null,
-  "local_data_directory_base": "/home/lennart_priorlabs_ai/code/large_data_ensemble/data-foundry/local-data-warehouse",
   "type_adapter_id": "dataset-mold-v1"
-}
+}