BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//talks.python-summit.ch//swiss-python-summit-2026//speak
 er//KZFYYW
BEGIN:VTIMEZONE
TZID:Europe/Zurich
BEGIN:DAYLIGHT
DTSTART:20251023T000000
TZNAME:CEST
TZOFFSETFROM:+0200
TZOFFSETTO:+0200
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251026T030000
RDATE:20261025T030000
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20260329T030000
RDATE:20270328T030000
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
SUMMARY:No more model training? TabPFN v3 vs. tuned trees. - Dominik Traxl
DTSTART;TZID=Europe/Zurich:20261023T135500
DTEND;TZID=Europe/Zurich:20261023T142500
DTSTAMP:20261111T040115Z
UID:pretalx-swiss-python-summit-2026-RCWDMG@talks.python-summit.ch
DESCRIPTION:What if you never had to train or tune a model again? That is 
 the pitch behind\ntabular foundation models: state-of-the-art predictions 
 with no training\, no\nhyperparameter search\, no feature engineering. Jus
 t call .fit() and .predict().\nThe headliner\, TabPFN\, whose earlier vers
 ion was published in Nature\, now adds\nclaims aimed straight at the gradi
 ent-boosted trees most of us reach for by\ndefault. TabPFN v3 promises to 
 ingest raw text natively\, scale toward a million\nrows\, support multi-cl
 ass classification natively and offer a\n"thinking mode" for extra quality
 . Too good to be true? I put it on trial.\n\nThis talk walks through a sys
 tematic\, reproducible Python benchmark that puts\nTabPFN\, exactly as it 
 ships\, up against a best-effort\, Optuna-tuned pipeline of\nXGBoost\, Lig
 htGBM\, and CatBoost on the messy public Steam Dataset 2025:\nnumerics\, b
 ooleans\, categoricals\, date-times\, free-text\, embeddings\, skewed\ndis
 tributions\, NaNs. The case spans three tasks: binary sentiment classifica
 tion\,\nmulti-class publisher prediction\, and Metacritic-score regression
 .\n\nWe stress-test each vendor claim one at a time: raw text versus sente
 nce\nembeddings versus no text\, out-of-the-box versus cleaned data\, defa
 ult versus\nthinking mode\, small samples up to large\, few to many classe
 s.\n\nBut the real takeaway is the method: how to benchmark a hyped claim 
 fairly\, why\nthe metric you pick can quietly decide the winner\, and the 
 data-leakage traps you\nmust defuse before you trust a single result.\n\nT
 he honest answer to "Too good to be true?" is: "it depends". And knowing w
 hat\nit depends on is the point. You'll leave with a feel for when a tabul
 ar\nfoundation model earns its keep\, when a tuned tree still wins\, and h
 ow to put\nthe next "no training required" promise on trial yourself.
LOCATION:Aula 4.101
URL:https://talks.python-summit.ch/swiss-python-summit-2026/talk/RCWDMG/
END:VEVENT
END:VCALENDAR