Account00054/o54_dataset_01_obfuscated

Dataset 01 Obfuscated

A tabular dataset with anonymized features and binned values, for training models that rank rows within each era.

Files

File	Description
dataset.parquet	The dataset: 1,742,712 rows and 1,145 columns
features.json	Feature sets, target names, and neutralizer features

Columns

Column	Description
id	Random row ID
era	Time order. A higher number is later
data_type	train, validation, holdout, or null in embargo gaps
eligible	Whether the row is in the scoring universe
tradable	Whether the row can be traded
target	The main target: 0, 0.25, 0.5, 0.75, or 1
target_*	32 other targets, on the same scale
feature_*	1,107 features, with bins from 0 to 4

Get started

1. Download the files:

   hf download Account00054/o54_dataset_01_obfuscated --repo-type dataset --local-dir o54_dataset_01_obfuscated
   

2. Load the training rows from the o54_dataset_01_obfuscated folder:

   import json
   
   import polars as pl
   
   info = json.load(open("features.json"))
   features = info["feature_sets"]["medium"]
   
   train = (
       pl.scan_parquet("dataset.parquet")
       .filter(pl.col("eligible") & (pl.col("data_type") == "train"))
       .select(["era", "target"] + features)
       .collect()
   )
   

Usage notes

• Score each era separately, then average across eras.
• Targets from nearby eras overlap. To score without overlap, use eras where (era - 1) % 24 == 0.
• Era numbers can have gaps.
• Don't use target columns as model inputs.
• Leave out holdout rows until you have a final model.