Train from public recipes
Architecture, data preparation, objectives, and compute assumptions stay in versioned source rather than disappearing behind a final metric.
Lab bootstrap · public surface live
Tabular foundation models,
trained in the open.
TabU-lab is a WeHub open research lab for building tabular foundation models from scratch. Every run begins with a falsifiable question and ends with enough provenance to reproduce, reject, or extend it.
- Open
- Code, configs, curves, mistakes
- Reproducible
- Fixed seeds, splits, compute
- Falsifiable
- Pass / kill gates before runs
PUBLIC TRAINING TRACE00 · BOOTSTRAP
typed table
compile→
hypothesiswritten first
executionseed + config + compute
verdictpending first run
Current truth
The lab infrastructure exists. No public training result, model checkpoint, or benchmark claim has been accepted yet.
See the live ledger →
01 · Mission
Make the path to a model inspectable.
“Open” means more than publishing a final checkpoint. The decisions that created it must remain visible enough for another researcher—or another agent—to audit and continue.
Be reproducible at small scale
Every idea first earns signal under an explicit budget. Scale is a later decision, not a substitute for understanding.
Publish failures as information
A killed hypothesis is a useful result when its protocol, logs, and boundary are clear. Silent dead ends teach nobody.
02 · Experiment protocol
Every experiment is a typed, auditable unit of work.
Marin shows the value of exposing the whole experiment lifecycle. TabU-lab adapts that idea to WeHub: a compact gate, an exact execution receipt, and an honest verdict before synthesis.
- 01
Preregister
Write the hypothesis, baseline, budget, and pass / kill criterion before execution.
- 02
Implement
Commit the code and config that make the proposed comparison executable.
- 03
Run
Record the command, seed, data provenance, environment, host, and compute used.
- 04
Receipt
Publish raw metrics, curves, logs, artifacts, and anything that limits interpretation.
- 05
Verdict
Mark pass, kill, or revise. Consolidate only what the receipt actually supports.
RUN RECEIPT / MINIMUM FIELDS
receipt.v0- hypothesis + gate
- commit + config hash
- dataset + split
- seed + environment
- host + compute budget
- raw metrics + curves
- failure notes
- pass / kill / revise
03 · Public ledger
The lab starts with an empty ledger—not invented momentum.
This page exposes the current state as it is. New cards appear only when a preregistration or receipt exists in the repository.
Public lab substrate
Repository layout, experiment receipt contract, and WeHub Research entrance are established.
Not preregistered yet
The first training question will be named here only after its hypothesis and gate are committed.
Stateno result claimed
04 · Research lanes
Four places to ask a falsifiable question.
These are invitation surfaces, not claims that the corresponding systems already exist.
Data & priors
Which synthetic or real-table distributions teach reusable structure without hiding test knowledge?
Representation
How should values, roles, missingness, units, and table structure become typed computational objects?
Training systems
Which small-scale recipes produce stable learning signals before expensive scaling?
Evaluation
Which held-out tables, shifts, and negative controls can actually falsify the proposed mechanism?
05 · Join the lab
Bring a question that can lose.
The best contribution is not a large promise. It is a sharp hypothesis, a matched comparison, and a receipt that lets the next person continue.