Spaces:
Running
Running
Commit History
Yesterday
-
Tour: button names show as pills, so the reader can match them to the screen Running
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>68988c6
2 days ago
-
Shortcut: "No leakage" rather than "No leakage warnings", so it does not read as hiding the warnings
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>66a29ce -
v0.1.9.12: leakage warning on LimiX-2 with recomputed pools, guided tour, benchmark groups in the rail
- LimiX-2 carries a leakage warning (autogluon/tabarena#648), read from the artifacts' new Leakage / LeakageDatasets columns: a badge and a red ! on its quality numbers in the Summary, no medals, a hover card with the evidence, and a mark in every chart. "Include methods with a leakage warning" (on by default) reads each pool's twin computed without it (entrants_<key>_noleak/, BeyondArena subsets_noleak/ and result_plots_noleak/); the API's records carry a leakage_warning field. - A guided tour (rail, Learn more, ?tour) of twelve steps, one of them on BeyondArena, and a once-per- browser note offering it to first-time visitors. - The rail lists TabArena and BeyondArena, and when opened the benchmarks built on TabArena (RamanBench) and those outside it (STRABLE and MulTaBench with their own logos, Time-series). - AutoGluon shows its own logo. - Data regenerated from tabarena's leakage-warnings branch (PR #691) plus its follow-up that publishes only the six pages per cell t…3cffdcc
4 days ago
-
Rename the resizer child for cached copies; HF-frame preview skill
Gradio serves iframe-resizer's child half without cache headers, so a browser that kept it from an earlier visit still ran its copy, which stretched the frame again and left the rail trailing the scroll. The app's page now asks for the file under another name, served empty like the original, so no cached copy is ever run. The preview harness is checked in as tools/hf_preview/ (a Hugging Face-like page framing the app), and the preview-in-hf-frame skill lists how to serve it next to the app, the frame checks (including a returning visitor), and the checks on the live Space after a push. CLAUDE.md and AGENTS.md point to both. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>658f725 -
Hero: "Why this benchmark?" opens the selling points
The band of four selling points returns, hidden at first: "Why this benchmark?" at the end of the hero's sentence opens it under the sentence, the numbers counting up as it opens, and a second click closes it. The Learn more pills keep the same numbers in their hover texts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>1c6376a -
Space frame scrolls inside itself; rail pinned there too
The app declines iframe-resizer: an ASGI middleware serves Gradio's child half as an empty script, and the head asks the Space page to let the frame scroll again (Gradio's page turns that off). The Hugging Face page then keeps the frame the size of the screen and the app scrolls inside it, as a static Space's does, so the rail and the dialogs are pinned by the browser instead of chasing the parent's scroll reports. For embedders that still size the frame to the whole app, the rail follows the visible band with a short glide over a strip the page paints down its full height. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>ae719ef -
Rail scrolls with the page in the Hugging Face page; v0.1.9.11
Inside the Hugging Face page the frame is as tall as the page and never scrolls, so a pinned rail could only follow the parent's scroll reports and trailed a beat behind. There the rail now sits at the top of the page, as tall as its links, and scrolls with it; on the Space's own address it stays pinned. The version history gains v0.1.9.11 for the redesigned pages. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>6c31c25 -
Site bar as a left rail on wide screens
From 861px the bar is a rail down the left edge: the icons alone, widening over the page on hover or keyboard focus with the labels fading in. It is pinned to the visible band like the dialogs, so it stays on screen inside the Space, and the page keeps 84px free on its left. Narrower screens keep the bar above the page. The figure "I care about" puts first no longer gets an accent frame; it still leads its tab. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>82ebe1f -
Benchmark pages as tabs under a compact hero
Each benchmark page shows one view at a time, picked from tabs at the top of its controls: Summary (the default), Leaderboard, Quality vs time, Head to head and Per dataset. Every view is on the page from the start, so a switch needs no round trip, and `?view=<key>` links to one. - The tabs, the one-row Customize bar and the shown view's first card form one block at full width. The controls sit on a surface a shade apart from the view, with a firmer line where it starts. - The tab captions moved to each tab's hover text, with what the selection covers. - The hero drops the "Meet our benchmark" eyebrow and the stats band; the numbers are in the Learn more pills' hover texts. - The floating bottom selection bar and its Jump to menu are gone. - The Summary table shows every row instead of scrolling inside a 680px box, and its heading takes the table's own surface. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>e3390a0 -
Summary table: Elo and improvability, in-cell bars, folded breakdowns
The Summary leads with Elo and improvability over all datasets and the two costs, as in-cell bars whose length is the value and whose colour runs from amber for the best to violet for the worst, each group closed by a divider. The breakdowns by task, size and target start folded into stub columns; one opens on a click, by itself when a selector points at it, and switches between Elo and improvability. A rank column leads the rows, the type column goes, the name colours and the marks (with a lock for API systems) are keyed in the Method header, which also holds the search, and the metric menu and family filters are gone. Fast end-to-end highlights both cost columns. The hero becomes one panel washed in the page colour under a Meet our benchmark eyebrow. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>573e3d0 -
Summary table: marks keyed in its headers, controls in its toolbar
The overview is titled Summary. What its marks mean moves from the heading's right into the headers of the columns that carry them (verified, non-commercial and the link under Method, the imputed star under the first number column, the dagger under the train time), and the top-3 key goes. The family legend, the metric menu, which way is better and the search move into a toolbar on top of the table, the legend on the left so the type key stays in view. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>f22ddb7 -
Customize your leaderboard starts closed
The selector card starts as a bar: its title is the toggle, shortcuts such as Add systems or Regression only pick a chip and open the card on it (a second click undoes them), and the column beside it keeps only the number of datasets the selection covers. The bar takes no wash of the page colour, so it no longer echoes the hero's band, and a link to the card (the bottom bar, Jump to) opens it. The Learn more pills and the dialog's tabs drop their emojis, and the agents topic is called MCP & API. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>935ad19 -
Benchmark your model ends the Learn more row
The call to submit a model moves from the hero's top links to the end of the Learn more row, washed lightly in the page's colour with an arrow for a link out, so it stands apart from the topics and draws the eye along the row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>14ec777
5 days ago
-
Update leaderboard data + version history (TabArena-v0.1.9.10)
The new leaderboard design ships with artifacts regenerated from tabarena main (d89e7962, after autogluon/tabarena#683): organization logos, explorers loaded from their files, the compact full table and the per-dataset browser beside its list. The version history gains v0.1.9.10 and links the archive of v0.1.9.9, the last version before the redesign. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>f9ad226 -
Dataset by dataset under the overview table
The per-dataset browser moves up from the end of the page to hang under the "All results at a glance" table, joined to it as the table's bottom bar: both look across the whole benchmark, unlike the leaderboard selected below. The bar takes the page's own colour, names how many datasets it covers, and opens from anywhere on it. On BeyondArena it sits under the overview figures. The floating selection bar now hides only while an open browser fills the view. The Follow control and its chips get a row of their own under "Your leaderboard", and the full table sits one tile gap under the figures: the two blocks drop the padding Gradio adds around every HTML block. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>a3355f5 -
Full leaderboard table as a card
The full table leaves its Gradio accordion for a card like the figure tiles: title, caption, and a button that opens it in the browser, with the download beside it once the table is up. Its frame loads only then. On TabArena the frame learns which methods the latest releases added, for its "new" badges, from the version history the overview table reads too. The Jump to entry for the section is named "Your leaderboard", like the section. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>04518bb -
Figure grid, follow a method, explorers from static URLs
The figures under "Your leaderboard" become one grid: the overview and the win-rate matrix span the width, the inference, end-to-end and train-time Pareto fronts and the tuning trajectories sit two to a row. "I care about" puts the figure that answers it first and "Measured by" picks each tile's metric, both in the browser from the reader's choice on the page element, so neither costs a render. Each tile has one header row (title, caption where it fits whole, actions) and an expand button that lifts it over the page. "Follow" in the leaderboard's heading, the follow button on each overview row, or a click on a method inside any explorer marks that method in every chart of the page; hovering a method in one explorer outlines it in the others. Explorers load from their artifact files (served as static paths, the URL versioned by the file's modification time) once they near the screen, with the site's theme in the URL, instead of travelling inline with every render: a first visit sends …dbafae3 -
Leaderboard section: spec line, flat figure panels
"Your leaderboard" and the per-dataset section get plain headings with one muted line under the title. For the leaderboard that line spells out the selection with the dataset count in bold: the task split, the size and target filters, who competes, imputation and TabArena-Lite. On BeyondArena it names the core2k protocol and warns when only a few datasets remain. What each section shows moves into the heading's hover hint, and the row of section links above the figures is gone (the Jump to menu covers them). Each figure is one flat card. Its header row holds the title, a caption that says what the figure shows, and two quiet buttons: Customize, which opens the interactive view and reads Done while it is open, and a Download menu with SVG, PDF and PNG. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>3e9a926 -
All results at a glance: logos, sortable table, end-to-end Pareto
The cross-subset table gets a plain heading, "All results at a glance", with the metric named in the line beside it and picked from a menu that says what each metric means, and the marks at the right. A row over the table filters it: a box to find a method and a toggle per family. A click on a number heading sorts by it, the sorted column's medal slot shows ranks below the podium, and a number's hover gives its rank in the column and, for Elo, its 95% interval. Rows keep their look and gain the organization's logo before the name, and a "new" badge for the methods the latest releases added. All columns fit at 1440px, the method column stays in view when the table scrolls, and on a phone the names shorten so numbers stay beside them. Logos come from the Organization column (MethodMetadata.organization upstream) and assets/logos/, served as static files; NOTICE notes they are their owners' trademarks. The Pareto front panel titled inference time opened on total time since the explorer …a977c5b -
Docs: Agents & API topic, dialog links, no appendix
The reference text at the bottom of the page moves into the Learn more dialog. The agentic guide becomes an Agents & API topic on both pages: how to choose a model automatically (TabArena only), the CSV paths and the MCP tools. The Appendix heading and its accordion are gone, and the Jump to menu opens the topic instead. BeyondArena's "What do the subsets mean?" box repeated the Subsets topic; it is now a button next to the overview figures that opens it. Every dialog has a link of its own: while one is open, ?doc=<name> is in the address bar and, inside the Space, in the Hugging Face page URL, and a link that carries it opens the dialog once its page shows. Code blocks in the dialog scroll inside their own box. An unused style rule and an unused script function are removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>1d28e06 -
Selector card: one card of chips with live dataset counts
Both pages pick their leaderboard in one card instead of Gradio tab bars and checkbox boxes. Each choice is a line of chips under a caption: task, datasets, who competes, what you care about with the headline metric, and the settings as switches. Every chip in a dataset row carries the number of datasets its leaderboard covers given the other choices, and a combination nobody published is greyed out instead of leading to an empty page. Size and target, and BeyondArena's dimensionality and feature type, are trays within one row; binary and multiclass sit in a tray right after classification. A click updates the card in the browser and sends the whole selection to the page's existing states in one event, so a quick run of clicks settles on the last. Beside the lines the card says what the selection covers: the number of datasets, set like the hero's, and a summary. The system warnings fold to one line each that opens on "Why?". The card carries its own title, so it follows the hero dire…772c0a0 -
Page heroes: one sentence, links, live selling points, docs dialogs
TabArena and BeyondArena open with the same hero, in the page's colour: one sentence on what the benchmark is, a Paper/Cite split button, a "Benchmark your model" link and a Team chip, then a band of four selling points. Their numbers are read from the data and the version history (datasets, models and systems, releases with the newest entrant, subsets), count up once and open the details behind them on click. The Team chip opens the maintainers with their GitHub avatars, the primary contact marked; BeyondArena lists its own paper's authors. The info pills and the About text give way to a "Learn more" row that opens one tabbed dialog with every topic in the page, so nothing waits on the server; the Metrics topic now defines each column. The TabArena page is called TabArena-v0.1, with the version set smaller in the top bar. Dialogs opened before the first scroll inside the Hugging Face page now fit the screen: the visible band is asked for again whenever the frame grows. Co-Authored-…b1c3a25
6 days ago
-
Top bar: benchmark links, version history and COI dialogs
Replace the centered title, Gradio's tab row and the floating Contents, COI and theme chips with one bar above the page tabs: the brand with its tagline, a button per benchmark (TabArena and BeyondArena framed in their own colour, the page on screen filled with it), and on the right Code, Data (Data Foundry), Our COI, the version chip and a light/dark toggle. The bar keeps one row while it fits and gives the page links a row of their own otherwise. The version chip and the version-history dialog are parsed from the existing version history text, so the accordion at the bottom of the page is gone. Both dialogs scroll inside themselves, so a click beside one always closes it. The Contents menu is now "Jump to" in the bottom selection bar, the one element left floating over the page. The Default/Light/Dark switch becomes a dark/light toggle; old ?theme= links still resolve. On phones the page gutter drops from 32px to 16px a side. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>33011b6
8 days ago
-
Update leaderboard data + version history (TabArena-v0.1.9.9)
dc73259
10 days ago
-
Update leaderboard data + version history (TabArena-v0.1.9.8)
Add the verified TabSTAR-v1.1 and remove TabSTAR (v1), generated from autogluon/tabarena#674 at 0aa07c95. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>e969363
16 days ago
-
Update leaderboard data + version history (TabArena-v0.1.9.7)
0bc60d4
18 days ago
-
Update leaderboard data + version history (TabArena-v0.1.9.6)
New verified models PrismBoost and LightPFN, and the first closed-source API system, Chakra-Tab, in its medium and full presets. Systems are colored by category, closed-source API systems in a rust shade, and the Type pill of the cross-subset overview carries the tag chip. A system that fits inside its prediction call counts that call as its train time, marked with a dagger. The closed-source API warning and the Systems panel say why such results should not be trusted by default. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>0aac376
20 days ago
-
Pinned chips: write the visible band onto them, not onto :root
Inside the Space, Hugging Face sizes the app's iframe to its whole content and scrolls the outer page, so the corner chips, the theme switch and the selection bar are re-pinned from the scroll reports iframe-resizer sends. Each report set --ta-vp-top / --ta-vp-h on :root, and a custom property on the root is inherited by every element, so every report restyled the whole page (13-22 ms) while the reader was scrolling. taSetBand writes the band onto the seven elements that read it instead (0.7 ms), and the scripts that need the numbers read them from taBand. Over one scroll gesture in an iframe-resizer embed, the app frame's style recalc drops from 514 ms to 31 ms; the chips land where they did. On its own *.hf.space address the Space gets Hugging Face's floating header injected into the top-right corner, over the theme switch; the switch now steps down below it there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VMfN1J8urroWZ…0c7a760 -
Theme switch in the top-right corner: Default, Light and Dark
Asked for in the Space's discussion #5: the dark page strains some readers' eyes over long use. The switch sits in the top-right corner, pinned like the Contents and COI chips. A click is remembered per browser, and a ?theme=light|dark|default link previews a theme without overwriting that choice. Default is the page as it was, pixel for pixel apart from the switch itself. Dark keeps the dark page and opens every figure in its dark edit view instead of paper view. Light is a soft light theme: a warm grey page, cards a step lighter, figures on their usual white paper, and the explorers' own light palette in edit view. All of it sits behind an html.ta-light class that only the switch sets. The heat cells get a light ramp of their own (views.HEAT_STOPS_LIGHT), written next to the dark shade as --ta-heat-light, which only the light theme reads. The full-table frame takes the same ramp through its srcdoc, and leaving Light gives every frame back the server's document. Co-Authored-By: Cla…149a7ab
23 days ago
-
Update leaderboard data + version history (TabArena-v0.1.9.5)
Add the verified Kumo-Tabular, Kumo-Tabular-Medium and Kumo-Tabular-Small (suites tabarena-2026-10-06 and beyondarena-2026-10-06) to TabArena and BeyondArena. BeyondArena moves from the 507-split core protocol to core2k (2,000 splits from the cost-weighted allocation): every BeyondArena cell is regenerated on core2k (autogluon/tabarena#647), and the page, API docstrings and Metrics & Protocol tab now name core2k. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>fa159b9
27 days ago
-
BeyondArena: combinable split / size / feature filters and a task selector (TabArena-v0.1.9.4)
The BeyondArena selector rows no longer reset each other: split, size, features and the new task row (classification, regression, binary, multiclass) each keep their own state, and every combination that holds at least one dataset is its own leaderboard (354 of 500, regenerated with autogluon/tabarena#634). An empty combination says so, and one with fewer than five datasets carries a warning. The per-dataset browser opens on the selected task, split and size. get_beyondarena_leaderboard takes split / size / features / tasks instead of subset, and get_pareto_frontier takes beyond_split / beyond_size / beyond_features, with tasks applying to both benchmarks. The CSV example in the TabArena info box points at the entrants_* layout. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>338a102
-
Update leaderboard data + version history (TabArena-v0.1.9.3)
TabArena: the Linear re-run with L1 honored for classification. BeyondArena: the nine foundation models of the 2026-09-22 run and the Linear re-run, in the new layout with TabArena's explorers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>0f16a23 -
BeyondArena page: TabArena's explorers below the overview figures, and ?tab= links to a page
The BeyondArena tab keeps its two cross-subset overview images and then uses the TabArena layout: a reduced control card (I care about, measured by, the subset axis as split / size / features chip bars, non-commercial toggle), the figure stack with the tuning trajectories, the win-rate explorer, the full table and the per-dataset browser opened on the selected subset. Every page tab can now be linked as ?tab=<slug>. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>34a2652
-
Update leaderboard data + version history (TabArena-v0.1.9.2)
2b3adbd
-
Add the Apache 2.0 LICENSE and a NOTICE naming the TabArena Maintainers
The Space README declared apache-2.0 but the repository carried neither the license text nor a notice. Both files match the TabArena repository. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>bc22c02 -
Point readers and the MCP guide to the per-split results export script
The per-dataset, per-split numbers behind the leaderboard are not hosted; say so in "Reading the numbers programmatically" and in list_leaderboards, and name examples/plots/run_export_results_per_split.py in the TabArena repository as the way to generate them. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>40edb65
-
Update leaderboard data + version history (TabArena-v0.1.9.1)
CatBoost returns to its July run (the re-run's fit times are not comparable yet), the per-dataset browser shows the Linear and KNN tuning trajectories again and marks methods without one, and the explorers keep a metric picked inside them across the paper-view toggle. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>9d41c20 -
Update leaderboard data + version history (TabArena-v0.1.9)
Re-run of the hosted foundation models and CatBoost under the timing and warm-up pipeline, TabPFN-3.5 and TabPFN-3.5-Fast (fit times re-measured), LimiX-2, the balanced / imbalanced / extreme dataset subsets, per-method licenses with the non-commercial toggle, and the explorers keeping a metric picked inside them when the host echoes its own. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>e24f3a7
-
An unpublished cell is one without a leaderboard table, not one without a folder
A skipped evaluation (a subset with no dataset of that task type) can leave an empty folder behind, so the notice keys on the table file every published cell has. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>6082baf -
Imbalanced means a 10:1 class ratio; an "Extreme imbalance" chip starts at 50:1
Copy and the Distribution chip bar follow the tabarena thresholds: imbalanced from a 10:1 largest-to-smallest class ratio (was 3:1), and a third chip for the extreme end at 50:1, a subset of the imbalanced datasets. The overview gets an Extreme column in its By Target group and the API accepts the value. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>6736d1c -
Log-scaled runtimes, a non-commercial toggle and a distribution row
Four leaderboard changes, matched by the tabarena branch of the same name. The cross-subset overview shades its Fit and Infer columns in log space, as the full table already did, so a 10x speed difference near the fast end is visible. The control card's "Protocol" row becomes "Settings" and gains "Include non-commercial methods". Off, methods whose `Commercial` column is false leave the overview, the embedded table and every explorer (told over postMessage via a `data-exclude` attribute on the frames). A row filter, deliberately unlike the entrant pools: the remaining rows keep their published numbers. The overview marks those methods with `$`, the license on hover. A "Distribution" chip bar (All | Balanced | Imbalanced) joins the dataset axis as a second bar next to the size buckets; the two share one state and reset each other. The overview gets a "By Target" group and the API accepts the two values. A cell not published yet renders a notice instead of failing. The per-dataset br…4767e1b
-
BeyondArena size subsets bucket datasets by their largest training split
Regenerated data_beyondarena with tabarena PR #530: the tiny, small, medium and large subsets were bucketed by the mean training split size, which put garments_worker_productivity in tiny and sf_permit_time in medium. They now follow the largest split, as the paper does, and hold 52 / 38 / 30 / 22 datasets. The other subsets are unchanged apart from regenerated figures. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>34272b6
-
Overview table gets a row per variant, and its cost columns follow the row
The cross-subset overview picked each column's value independently, taking the extremum per model. For the two cost columns that meant the minimum, which is always the default variant, while the model cell was labelled with whichever variant led on the selected metric. A row reading "RealMLP (tuned + ensembled)" quoted the default's fit time: 10.06 against 2950.72, roughly 200x low, because tuning fits 200 configurations where the default fits one. 20 of 39 model rows were affected, by 6.8x to 902x. Rows are now keyed by (model, variant), so default, tuned and tuned + ensembled each get their own line and a row's cost is the cost of the run it scores. That removes the mismatch by construction, and puts the price of tuning on the page. "One per model" collapses each method to its best variant, decided by the leading column. It sits in the overview's own row and feeds only that table, and is named after the win-rate matrix's button so the two read as one control. The upstream artifact…17d9c9b
-
Update leaderboard data + version history for ChimeraBoost 0.30.0
9b53220
-
Per-dataset results leave out the models that could not run on a dataset
An imputed score stands in for a model that could not run on that dataset, so the per-dataset view no longer shows it there -- the model keeps its place on every dataset where it did run. The size of the field now varies by dataset, and the pane says how many could not run. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>74c6c71 -
Selection bar stays down past the per-dataset section
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>dcf9e02 -
Charts no longer draw themselves before they can be measured
A chart drawn while its container was still unlaid-out pinned itself to its minimum width and stacked its labels, and nothing later healed it. It now waits for a measurable box, never caches a zero text measurement, and watches its own box rather than the viewport. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>5cf9f1a -
Systems as their own class of entrant, and per-dataset results
Systems (AutoGluon, TabFM+, agents, hosted APIs) are now their own kind of entrant rather than a reference pipeline alongside the models. Every leaderboard number is relative to the field, so each combination of system categories is its own published evaluation rather than a row filter, picked with the new "Who's competing?" control; models-only is the default and is now genuinely models-only. New "Per-dataset results" section: the benchmark one dataset at a time. Pick a contender and each row reports its rank and its signed margin on that dataset, with a star placing it in the field; select a dataset for its own tuning trajectories and its full ranking. Also: an "I care about" control that leads with the figure answering your question, the win-rate matrix carrying every tuning variant again, and the pinned overlays staying put inside the Space. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>9a668d8
-
Refresh leaderboard data with the re-run EXAONE-Tabular results
EXAONE-Tabular now covers regression as well as classification, so its regression numbers are measured instead of imputed. Its classification runs were redone on the updated released checkpoint and on the classifier default n_svd=0 that its authors published, which moves those numbers too. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CJAJDEkj2fiLTtjeLo5APw3e8e748 -
Revert "Refresh leaderboard data with the re-run EXAONE-Tabular results"
This reverts commit d96dd9846bbc2f9caffd31fd20055e720e08f98e.32fa5a1