A self-contained, sortable comparison page of every TabArena entry that scores at least as high as TabICLv2, with model facts (architecture, pretraining, synthetic priors, ICL details, limits, speed/cost, licensing) compiled from the papers, repositories and model cards, plus TabArena and TALENT benchmark numbers.
Open index.html in a browser (works from file://, no server needed), or serve the folder:
python3 -m http.server 8765 --directory /home/anne/Projects/tfm_comparisonindex.html the page (HTML + CSS + JS, no external dependencies; reads data/compiled.js)
data/
SCHEMA.md field-by-field definition of a model record
models/<id>.json one hand-curated record per model (the thing to edit / extend)
tabarena.json parsed leaderboard (all subsets, win-rate matrix, licenses) — generated
talent.json TALENT results compiled per source table (protocol-aware)
compiled.js / .json merged data consumed by index.html — generated, do not edit
raw/ leaderboard snapshots (CSV export from the TabArena Space, page text, version history)
scripts/
fetch_tabarena.sh download the current leaderboard CSVs and rebuild data/tabarena.json
build_tabarena.py CSV -> data/tabarena.json
build.py models + tabarena + talent -> data/compiled.js
- Create
data/models/<id>.jsonfollowingdata/SCHEMA.md(copy an existing record as a template). Usenullfor unknown facts; numeric fields used for sorting must be plain numbers. - Set
tabarena_entryto the exact row label on the leaderboard (e.g."NewModel (default)"). - Refresh the leaderboard snapshot if needed:
scripts/fetch_tabarena.sh(then update theCurrent Version:line indata/raw/tabarena_version_history.md). - Add TALENT numbers to
data/talent.json(one table per source + protocol; results keyed by model id). - Rebuild:
python3 scripts/build.py— it prints warnings for records that do not join to the leaderboard. - Reload
index.html.
Columns in the table are generated from the column config at the top of the script in index.html
(GROUPS); the expanded fact sheet renders every field in the record generically, so new fields appear
without touching the page.
A record may carry an optional tabarena_self_reported block (see data/SCHEMA.md) with the authors' own
TabArena numbers. The page shows them in the TabArena columns marked with † and excludes the model from the
win-rate matrix and the timing scatter; as soon as scripts/fetch_tabarena.sh brings in an official row with the
same tabarena_entry label, the official numbers take over automatically (the build prints a note).
Previous example: NVIDIA Kumo Tabular (provisional from 2026-09-30 until its official rows appeared, 2026-10-07). No model is provisional at the moment.
TabArena leaderboard, "Models only" view, all tasks, all datasets, all repeats, imputed and non-commercial methods included: every entry with Elo >= TabICLv2 (default). RealTabPFN-2.5 qualifies through its "tuned + ensembled" variant only.