<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Getting Started on tab-foundry Docs</title><link>https://bensonlee5.github.io/tab-foundry/docs/getting-started/</link><description>Recent content in Getting Started on tab-foundry Docs</description><generator>Hugo</generator><language>en</language><copyright>tab-foundry contributors</copyright><atom:link href="https://bensonlee5.github.io/tab-foundry/docs/getting-started/index.xml" rel="self" type="application/rss+xml"/><item><title>tab-foundry</title><link>https://bensonlee5.github.io/tab-foundry/docs/getting-started/repo-overview/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://bensonlee5.github.io/tab-foundry/docs/getting-started/repo-overview/</guid><description>&lt;p&gt;A tabular foundation model that generates the data it learns from.&lt;/p&gt;
&lt;p&gt;&lt;a href="https://github.com/bensonlee5/tab-foundry/blob/main/LICENSE"&gt;&lt;img src="https://img.shields.io/badge/License-Apache_2.0-blue.svg" alt="License: Apache-2.0"&gt;&lt;/a&gt;
&lt;a href="https://www.python.org/downloads/"&gt;&lt;img src="https://img.shields.io/badge/python-3.14-blue.svg" alt="Python 3.14"&gt;&lt;/a&gt;
&lt;a href="https://bensonlee5.github.io/tab-foundry/"&gt;&lt;img src="https://img.shields.io/badge/docs-bensonlee5.github.io%2Ftab--foundry-blue" alt="Docs"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Most tabular foundation models learn from a fixed corpus and stop. You get a
&lt;code&gt;.predict()&lt;/code&gt; call and a benchmark number, but no control over the data the
model trained on, the architecture it uses, or the training loop that produced
it.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;tab-foundry&lt;/strong&gt; takes a different approach. It uses
&lt;a href="https://github.com/bensonlee5/dagzoo"&gt;dagzoo&lt;/a&gt; to generate synthetic tabular
datasets, trains them with an active sandwich lane plus a frozen PFN control
and historical staged reference lane, benchmarks against real-world tasks, and
exports inference bundles you can deploy. You control the full pipeline: what
data gets generated, which model surface is active, how training runs, and
what gets exported.&lt;/p&gt;</description></item><item><title>Problem Formulation</title><link>https://bensonlee5.github.io/tab-foundry/docs/getting-started/problem-formulation/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://bensonlee5.github.io/tab-foundry/docs/getting-started/problem-formulation/</guid><description>&lt;p&gt;The primary objects are:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;data-side prior parameters &lt;span class="math-inline" data-katex-source="\phi"&gt;\phi&lt;/span&gt; for the synthetic task generator&lt;/li&gt;
&lt;li&gt;model-side parameters &lt;span class="math-inline" data-katex-source="\theta"&gt;\theta&lt;/span&gt; for the sandwich model&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Everything else in the repo should be understood as machinery for sampling
from the induced task distribution and optimizing the resulting objectives.&lt;/p&gt;
&lt;p&gt;Use these alongside this page:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;architecture reference:
&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/development/model-architecture/"&gt;Model Architecture&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;canonical roadmap:
&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/development/roadmap/"&gt;Roadmap&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;workflow runbooks:
&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/ml-engineering/workflows/"&gt;Workflows&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;sweep contract:
&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/research-contributors/sweep-contract/"&gt;Sweep Contract&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;reference index:
&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/reference/"&gt;References&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="problem-setup"&gt;Problem Setup&lt;/h2&gt;
&lt;p&gt;Let a supervised tabular task be&lt;/p&gt;
&lt;div class="math-display" data-katex-source="T = (X_{\mathrm{tr}}, Y_{\mathrm{tr}}, X_{\mathrm{te}}, Y_{\mathrm{te}}, \tau)."&gt;
T = (X_{\mathrm{tr}}, Y_{\mathrm{tr}}, X_{\mathrm{te}}, Y_{\mathrm{te}}, \tau).
&lt;/div&gt;
&lt;p&gt;with:&lt;/p&gt;</description></item><item><title>Glossary</title><link>https://bensonlee5.github.io/tab-foundry/docs/getting-started/glossary/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://bensonlee5.github.io/tab-foundry/docs/getting-started/glossary/</guid><description>&lt;p&gt;Use this glossary when architecture, sweep, artifact, and workflow terms start
to carry too much repo-specific meaning.&lt;/p&gt;
&lt;h2 id="anchor"&gt;Anchor&lt;/h2&gt;
&lt;p&gt;The locked comparison run for a sweep. New rows are judged against the anchor
unless the queue row explicitly declares a different preserved surface.&lt;/p&gt;
&lt;h2 id="architecture-screen-surface"&gt;Architecture-Screen Surface&lt;/h2&gt;
&lt;p&gt;The canonical benchmark-facing sandwich surface used for current architecture
work. Historical staged surfaces still exist for comparison, but they are no
longer the default landing zone for new architecture evidence.&lt;/p&gt;</description></item><item><title>Contributing</title><link>https://bensonlee5.github.io/tab-foundry/docs/getting-started/contributing/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://bensonlee5.github.io/tab-foundry/docs/getting-started/contributing/</guid><description>&lt;p&gt;Use this guide when you want to make a bounded change without reopening the
entire system.&lt;/p&gt;
&lt;p&gt;Start with &lt;a href="https://bensonlee5.github.io/tab-foundry/docs/getting-started/repo-overview/"&gt;README.md&lt;/a&gt;, then go directly to the owner docs that
match your question:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/getting-started/repo-overview/"&gt;README.md&lt;/a&gt; for the repo overview and quickstart&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/ml-engineering/workflows/"&gt;docs/workflows.md&lt;/a&gt; for command examples and artifact
expectations&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/research-contributors/sweep-contract/"&gt;program.md&lt;/a&gt; for the active system-delta sweep contract&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/development/roadmap/"&gt;docs/development/roadmap.md&lt;/a&gt; for research
priorities and TF-RD sequencing&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/development/model-architecture/"&gt;docs/development/model-architecture.md&lt;/a&gt;
for the live model surface&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/development/codebase-navigation/"&gt;docs/development/codebase-navigation.md&lt;/a&gt;
for package ownership and entrypoint boundaries&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/ml-engineering/inference/"&gt;docs/inference.md&lt;/a&gt; for export/runtime handoff details&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bensonlee5.github.io/tab-foundry/docs/getting-started/glossary/"&gt;docs/glossary.md&lt;/a&gt; for sweep and architecture vocabulary&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;If a Markdown file and &lt;code&gt;.venv/bin/tab-foundry ... --help&lt;/code&gt; disagree, trust the
CLI and update the Markdown file.&lt;/p&gt;</description></item></channel></rss>