Skip to content

docs(benchmarks): focus leaderboard on aggregate results - #194

Merged
Sonorous281 merged 14 commits into
RLinf:mainfrom
ye1457:docs/simplify-benchmark-leaderboard
Sep 17, 2026
Merged

Sonorous281 merged 14 commits into
RLinf:mainfrom
ye1457:docs/simplify-benchmark-leaderboard

Conversation

@ye1457

@ye1457 ye1457 commented Sep 16, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Latest review follow-up: pin the Back to top anchor fix from misc fix: align EGL device with --cuda-device to prevent env_server crash on multi-GPU #62 and add the same fire icon as adjacent What's New entries in both READMEs. Fast-forwarded the remote PR head to preserve the maintainer's merge of Flash Mode feat(flash): introduce evaluation-only Flash Mode with selected memory #195; no runtime edits or result changes in this follow-up.

  • Review follow-up: use theme-aware misc assets and one jsDelivr commit for normal benchmark resource loading. Correct the bilingual native fallback grouping, place LIBERO-PRO and RoboCasa matrices under their own headings, and link README to the short Astra memory-batch explanation (Long 157/200 + other suites 584/600 = 741/800). No score or denominator changes; no public theme or runtime changes.

  • Sync the contributor's latest request: remove the RoboCasa Astra seed/episode note from both native tables and pin the matching bilingual misc assets. Scores, episode counts and CSV metadata remain unchanged.

  • Latest presentation update: ASPIRE uses a plain method label, with only the Long Task / Long Swap zero-shot note below the table. Values and underlying calculation metadata remain unchanged. All eight Astra LIBERO-PRO components are present; RoboCasa Astra split scores are now complete: Atomic 87.78% (79/90), Seen 43.75% (35/80), Unseen 42.50% (34/80), Overall 59.20% (148/250).

  • September 17 update: synchronize text-only leaderboard assets and max.reasoning labels for Opus-4.7 in both languages, including native no-JavaScript tables. Scores and episode counts are unchanged.

Ready for review at the contributor's request. Resource PR: RLinf/misc#62.

Based on main after #193 reverted #165. Add a compact benchmark page, native fallback summaries and navigation below Quick Start, without restoring the previous large provenance, per-task, paper-analysis, zero-shot or demo sections. Existing runtime, tutorial navigation and technical Task Card usage docs remain unchanged.

  • Pin misc resources to c3b9b5d4ffa360a8324c5b7aa510e1ed0876aa43; all implementation, result data and media remain in misc.
  • Rename the existing complete LIBERO-PRO result to RPent Flash Mode, 72.63% (581/800); Molmo2-8B is only visual localization. Remove the old 200-episode subleaderboard.
  • Rank ASPIRE's 61.36% arithmetic mean of its eight displayed scores at the contributor's request, with explicit derived-score metadata and the Long zero-shot scope retained. Remove the separate Flash summary module; fill its eight independent Task/Swap cells from 800 audited contributor cases: Spatial 79% / 70%, Object 86% / 93%, Goal 74% / 65%, Long 60% / 54%. No merged columns or substituted aggregate values.
  • Remove NORA/OpenVLA only from LIBERO-PRO. Native bilingual fallback tables include the new detailed scores and scope notes.
  • Include Xiaomi-Robotics-1's official 57.4 / 80.2 / 57.1 / 32.1% Overall/Atomic/Seen/Unseen scores. Do not assign RPent's 340-episode denominator or assert verified protocol equivalence.
  • Restore RoboTwin GPT-5.5 xhigh 62.4% (156/250) with its published source; GPT-6/GPT-5.6 stay omitted.
  • Preserve all other previously approved model corrections and Astra 92.63% (741/800).
  • Keep bilingual native tables for JavaScript/remote-resource failures. No code-directory README, implementation code or media is added to RPent.

Validation

  • Back-to-top follow-up: 112 focused pointer/keyboard checks passed across both languages, themes and viewport sizes, including standalone and offline pages. Both README announcement lines now use the same fire icon as adjacent entries. Maintainer merge a5442c4 is preserved.

  • 284 additional review checks passed for both themes/languages and desktop/mobile, system/document theme changes, README deep links, themed controls, no-JavaScript and resource-failure fallback, and offline documents. Original 629 browser checks and strict bilingual builds pass.

  • The supplied RoboCasa Astra archive contains 50 tasks × 5 seeds = 250 cases, not 340. Correct its tooltip/CSV denominator to 148/250; include all three split scores in interactive and bilingual native tables. Other models and the public 340-cell runtime protocol remain unchanged. No raw archive is committed.

  • 629 Playwright checks passed, covering 19 views, both languages, desktop/mobile, exact scores, independent-column sorting, CSV, keyboard, no-JavaScript fallback and offline preview.

  • Data audit: 171 records; unknown counts stay empty, Flash Overall appears once, no separate Flash summary remains, ASPIRE's mean is explicitly derived.

  • Strict clean English/Chinese Sphinx builds and pre-commit pass.

  • No runtime code changed. Prior CPU regression was 593 passed, 3 skipped; this update reran presentation tests and source-sensitive scans.

  • Both languages loaded the immutable public misc commit successfully and matched the expected scores. Existing Sphinx theme-mode console warnings are unrelated to the benchmark renderer.

Merge misc #62 first; no automatic merge requested.

@ye1457
ye1457 marked this pull request as draft September 16, 2026 12:57
@ye1457
ye1457 force-pushed the docs/simplify-benchmark-leaderboard branch from ee6cea8 to 67014d8 Compare September 16, 2026 12:57
@read-the-docs-community

read-the-docs-community Bot commented Sep 16, 2026

Copy link
Copy Markdown

@read-the-docs-community

read-the-docs-community Bot commented Sep 16, 2026

Copy link
Copy Markdown

@ye1457

ye1457 commented Sep 16, 2026

Copy link
Copy Markdown
Contributor Author

Applied the latest contributor corrections in c19727c with misc snapshot 7551b6b: RoboTwin removes GPT-6/GPT-5.6, restores GPT-5.5 58.0%, and uses Opus-4.7; standard LIBERO and RoboCasa also use Opus-4.7. LIBERO-PRO removes Opus-4.8 and both GPT-5.6 configurations, retaining Opus-4.7 Overall 82.4% without reassigning the removed model’s suite scores. Chinese/English fallback tables, CSV, overview images and local offline previews are synchronized. 303 browser checks and strict bilingual builds pass; actual immutable CDN rendering was verified in both languages. Keeping this PR and misc #62 Draft pending Chinese-preview approval.

@ye1457
ye1457 marked this pull request as ready for review September 16, 2026 16:00
@qurakchin

Copy link
Copy Markdown
Collaborator

please fit dark mode
image

Comment thread docs/source-en/rst_source/benchmarks.rst Outdated
Comment thread docs/source-en/rst_source/benchmarks.rst Outdated
Comment thread README.md Outdated
@ye1457

ye1457 commented Sep 17, 2026

Copy link
Copy Markdown
Contributor Author

Addressed the dark-mode request in ed2f896 with misc 4442039. The embedded leaderboard now follows the live Sphinx theme, including the existing theme-toggle button; standalone/offline pages follow the system preference. Backgrounds, text, tables, menus, tooltips and action icons are themed without changing the shared documentation theme. Overview figures were regenerated using the actual light/dark renderer. Validation: 629 existing browser checks plus 284 review checks passed, strict English/Chinese builds passed, and both languages loaded the public immutable CDN resources and switched themes successfully. Existing upstream theme initialization warnings remain recorded separately. Preview: English / Chinese. Both PRs remain Ready for review; no discussions were marked resolved.

Comment thread docs/source-en/rst_source/benchmarks.rst Outdated
Comment thread docs/source-zh/rst_source/benchmarks.rst Outdated
@Sonorous281
Sonorous281 merged commit 180a9be into RLinf:main Sep 17, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants