ROGII - Wellbore Geology Prediction (Kaggle Competition Data)

It is the raw material of a three-month contest that turned geosteering into a leaderboard sport. ROGII, a directional-drilling software company, posed a deceptively simple question: given a drilled horizontal well and its vertical reference, can you predict the geology encountered along the lateral? The prize pool was $50,000 ($25k/$13k/$7k/$5k plus optional working-note awards). The response was loud: 16,341 entrants, 6,956 participants, 6,125 teams filing 161,975 submissions between May 5 and August 5, 2026.

The framing matters because it tells you why the data is shaped the way it is. ROGII estimates roughly 10,000 horizontal wells are drilled yearly, and each one requires someone to interpret - foot by foot - where the bit sits relative to the formation it is supposed to stay inside. Get it wrong and you drill out of zone: wasted footage, missed pay, geological hazards. Automating that interpretation is the product direction, and this dataset is its training problem stated publicly.

What arrives is compact and dense: 1.33 GB across 2,327 files and roughly 12,400 columns, organized as train and test sets keyed by 8-character well hashes. Every visible training well carries three files - lateral CSV, typewell CSV, PNG cross-section - and about 200 test wells hold their TVT back. A 28.79 MB task deck documenting the objective ships alongside.

In Datadory's catalog of 1,744 datasets across 159 viable industries, this record anchors the ML-benchmark end of the fourteen-dataset oil and gas equipment and services set. Amaze yourself with what's in it, then get the rows that matter:

Get a sample of this dataset

Structure first, because that is what a well hash conceals. One training well, 015fe0d2, unpacks into three files:

015fe0d2__horizontal_well.csv
  WELLNAME, MD, X, Y, Z, ANCC, ASTNU, ASTNL,
  EGFDU, EGFDL, BUDA, TVT, GR, TVT_input

015fe0d2__typewell.csv
  TVT, GR, Geology

cross-section.png   <- the interpretation, drawn

Read the pairing and the design snaps into focus. The horizontal well file walks the lateral one station per foot of measured depth, each station positioned in space (X/Y/Z) and instrumented with gamma ray. The typewell file is the same story told vertically: a TVT index, its matching GR signature, and a categorical Geology label naming which formation each interval belongs to. Correlating the second against the first is the classic geosteering workflow - and the six horizon columns (ANCC through BUDA) pre-load one machine-generated guess of where named formations sit along the lateral.

The submission format closes the loop: rows keyed {WELLNAME}_{row_index} - 015fe0d2_1654,-38.7 in the published example - with the predicted TVT scored by root mean squared error. That single row is the entire deliverable, which is why the label column is the dataset's real asset.

Sample rows above show file-level shape observed during the August 2026 research pass; row-level readings ship inside your sample.

Every column in the bundle has a verified definition, which puts this record ahead of most drilling data you will meet. The geometry spine comes first: MD measures depth in feet down the hole; X and Y give easting and northing position; Z gives true vertical depth below sea level. Together they reconstruct the trajectory in three dimensions, so build rates and lateral reach are derivable rather than asserted.

GR is the star witness: gamma ray in API units, present in both the lateral and the typewell files, measuring natural radioactivity so shale intervals read hot and clean sands read quiet. It is the channel that makes correlation between vertical reference and horizontal well possible at all.

The six horizon columns - ANCC, ASTNU, ASTNL, EGFDU, EGFDL, BUDA - carry predicted depths for named formations along the lateral, training files only. Treat them as one interpreter's prior, not truth: the whole point of the exercise is that the manual interpretation supersedes them.

That interpretation is TVT, true vertical thickness in feet - the manually interpreted geological position recorded once per foot of lateral, and the competition target. Its twin TVT_input repeats TVT as a model input with the evaluation zone masked to NaN, a deliberate leak-control device worth copying in your own pipelines. On the typewell side, Geology labels each vertical interval with its formation code, and the submission schema defines id/tvt for scoring.

Definitions above are verified against the dataset's own documentation.

Geography: deliberately obscured. Wells are anonymized to 8-character hashes, positions ride as X/Y easting-northing without coordinate-reference metadata in the public description, and no basin is named. If your models must generalize across basins, that anonymity is a feature - but pin down the CRS on your sample before mapping anything, because projected coordinates without a declared projection are an ambiguity waiting to bite.

Temporal: a single static snapshot, released for the competition window of May 5 through August 5, 2026 and frozen when it closed. There is no time series here and no update cadence to monitor - this is a benchmark, not a feed. Pair it with a daily-refreshing regulator if your pipeline needs movement.

Granularity: the finest cut in commercial well data - one sample per foot of lateral, each foot carrying geometry, gamma ray and (in training) a human-interpreted TVT value. Roughly 200 visible test wells mirror the training structure with TVT withheld in the evaluation zone; the visible test folder itself holds only a handful of authoring examples, swapped for a hidden set at rerun. Three files per training well, 2,327 files total, and the arithmetic works out to roughly 12,400 columns across the bundle.

API, files, or your warehouse. Daily, weekly, or hourly.

Name the wells or the shape you want - full laterals, typewell pairs only, or a slice cut to specific formation intervals - and the sample is assembled around that. Because each well is self-contained under its hash key, samples cut cleanly along well boundaries with no join risk.

Two practical notes before you build. First, the files are plain CSVs with a fixed column order, so they load identically in pandas, Polars, DuckDB or a warehouse loader - no conversion layer, no proprietary formats, and the PNG cross-sections travel alongside as documentation of the interpretation rather than as data to parse.

Second, respect the masked column. TVT_input blanks the evaluation zone precisely so models cannot cheat; if you re-split the data for your own validation, reproduce that masking discipline or your metrics will flatter you. Local validation should lean on splits of the training wells, since the held-out wells' answers are not inspectable.

If you want the horizon codes mapped onto your own formation vocabulary, say so when requesting the sample - that mapping is the usual first cut we make.

• Geosteering model development - per-foot TVT labels across ~200 laterals are the supervised target that well-log archives almost never include; train where the interpretation already exists instead of paying geologists to create it one well at a time.
• Log correlation automation - every lateral ships beside its vertical typewell, so GR-to-GR matching becomes a testable algorithm with a drawn answer sheet attached.
• Benchmarking - 161,975 scored submissions under identical RMSE rules give any new method a public baseline; beating the median of 6,125 teams means something.
• Competitive intelligence - the task deck inside the bundle states how a directional-drilling vendor frames automated interpretation, which is market research money cannot otherwise buy.

Each job maps onto a persona page: data scientists, developers builders, and investors quants reading the participation numbers as demand signal.

Data scientists and ML engineers lead, because labeled subsurface interpretation is the binding constraint on drilling-side ML everywhere - sensor streams exist in abundance, interpretations do not.

Developers and data-product builders come second: hashed well keys, fixed column order and a demonstrated row-identity convention mean the bundle integrates without a cleanup project.

Competitive intelligence and product teams read the competition itself - prize structure, task framing, participation curve - as a document about where drilling software vendors see automation heading. Investors and quants read the same numbers the same way: $50,000 bought three months of 16,341 entrants, which is a demand measurement disguised as a contest. Academics and students get a benchmark with explicit rules and a public leaderboard, the honest way to teach geosteering ML.

This category is full of activity signals and infrastructure maps - rig counts, incident statistics, platform detections, compressor sensors. This record supplies the thing none of those have: a supervised geological target, interpreted by humans, once per foot, ready to train against.

Against OPD's Sentinel-1-derived offshore platform dataset, the contrast is detection versus interpretation: radar-derived vessel and structure observations at basin scale, versus per-foot geological truth on a few hundred wells. Against live sensor feeds like the Valhall compressor stream, this is episodic and labeled where those are continuous and unlabelled. Use them together - benchmarks to teach the model, feeds to run it.

The direct comparison is written up in OPD vs ROGII Wellbore Geology Prediction, and the wider ranking lives at best oil gas equipment services datasets.

Know which question you are asking of the data. If it is geosteering model training, ask for training wells complete with their typewells and horizon columns; if it is correlation-tool prototyping, the typewell-lateral pairs matter more than volume.

Know also the honest boundaries documented during the August 2026 research pass: geographic provenance is undisclosed beneath the hashing, exact well counts beyond 'about 200' visible test wells are unconfirmed, and the size reported for the bundle (1.33 GB rendered) differs from the figure cited in the assignment text (~10 GB), so scope is confirmed at sampling rather than promised blanket here.

Finally, know your destination. CSVs land anywhere, but if your stack expects Parquet or database tables, mention it and the sample arrives converted - the per-foot grain multiplies rows quickly, and a full lateral is a different storage proposition than a summary table.