Commit History
-
Use dense ranks for leaderboard scoring Running
MedGRPO Teamf35d192
-
Add ECCV'26 legend to official leaderboard footnote
MedGRPO Team +1: Claude Fable 5Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>cf4e327 -
Show ECCV'26 badge on official leaderboard entries with is_eccv
MedGRPO Team +1: Claude Fable 5Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>363469e
-
Fix invisible ECCV badge text in leaderboard table
MedGRPO Team +1: Claude Opus 4.8 (1M context)The .lb-table td * { color:inherit !important; } rule forced the in-table ECCV badge text to take the cell color, overriding its inline gold color and making "ECCV'26" blend into the dark badge background. Add an .eccv-badge exemption so it renders gold-on-dark consistently with the footnote badge. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>f6a4e12
-
Clarify LLM Judge is unavailable in Evaluate Only mode
MedGRPO Team +1: Claude Opus 4.7 (1M context)Step 1 already runs with --skip-llm-judge, and Step 2 (LLM Judge) requires a published leaderboard row to look up the model and write scores back — neither of which exist in Evaluate Only. So DVC_llm / VS_llm / RC_llm will always read as 0.0 in eval-only runs. Surface this so users expecting captioning scores don't get confused: - Inline info text on the Evaluate Only checkbox. - Blockquoted "Heads up" note in the eval-only success message. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>3bf5e65 -
Add Evaluate Only mode to avoid leaderboard noise
MedGRPO Team +1: Claude Opus 4.7 (1M context)Addresses community feedback (Space discussion #1): submitters who only want to know their numbers were forced to publish, producing duplicate "test/test2/test3" rows on the public table. When the new "Evaluate Only" checkbox is ticked: - Validation + full evaluation run as normal. - A would-be rank is computed against a temporary copy of the leaderboard. - Nothing is saved: save_leaderboard and backup_results_to_repo are skipped. - The duplicate-name check is bypassed so users can iterate freely. ECCV 2026 Challenge entries are not allowed to use Evaluate Only (challenge submissions must be published). The form disables and un-checks the box automatically when the ECCV checkbox is ticked, and the server enforces the same rule defensively. Drive-by: removed an early "Please provide the authors" check that rejected community submissions even though the Authors field is hidden in non-ECCV mode. ECCV-required validation still catches the missing authors case for challenge entri…892e0a6 -
Drop the Remark field
MedGRPO Team +1: Claude Opus 4.7 (1M context)Remark was a free-form note that never appeared on the leaderboard and played no role in validation, ranking, or display. The is_eccv boolean column (and the matching badge) already distinguishes challenge entries from community ones, making Remark redundant. Removed from the form, submit_model signature, persisted entry, success message, and the empty- leaderboard schema. Existing rows with a "remark" key continue to load — pandas just ignores the extra column on read. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>ec95c47 -
Require Authors for ECCV 2026 Challenge submissions
MedGRPO Team +1: Claude Opus 4.7 (1M context)Authors is now mandatory alongside Team Name and Email for is_eccv=True submissions. Form labels and the checkbox info text updated to reflect the full required set. Community (non-ECCV) submissions are unchanged — Authors is still hidden in that mode. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>b3ed5db -
Hide ECCV-only fields from non-challenge submitters
MedGRPO Team +1: Claude Opus 4.7 (1M context)Team Name, Authors, and Remark start hidden and appear only when the ECCV checkbox is ticked. Remark auto-fills with "ECCV Challenge 2026" on the same toggle. Cleans up the form for the common case (community uploads) while keeping all challenge-required inputs one click away. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>714aa1a -
Fix ECCV banner text contrast on dark background
MedGRPO Team +1: Claude Opus 4.7 (1M context)The theme's global heading color was overriding the inline white text via specificity, leaving the "Now hosting the…" line dark navy on navy and nearly unreadable. Force white/cream on banner text and link with !important so the theme stylesheet can't override it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>0c636f5 -
Add MedVidU @ ECCV 2026 Challenge support
MedGRPO Team +1: Claude Opus 4.7 (1M context)Submission form: - New "ECCV 2026 Challenge" checkbox; Team Name and Email become required when ticked, optional otherwise. - Add Team Name, Authors, Remark fields; rename Organization to Institution; replace Contact with Email. Persist team_name/authors/remark/is_eccv in leaderboard rows. Leaderboard page: - Top banner above the tabs announcing the workshop with a link to https://medvidu.github.io/ and a pointer to the Submit tab checkbox. - Small "ECCV'26" badge next to model names with is_eccv=True; legend entry links to the workshop. Existing rows (is_eccv missing/NaN) are rendered unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>93842b2
-
Each leaderboard ranks within itself; no rank gaps
MedGRPO Team +1: Claude Opus 4.7 (1M context)The previous global-rank scheme (commit 03adf33) made the same model show the same rank number in both Official and Community tables by ranking against the union of all submissions. Side effect: the Official table displayed ranks like 2, 3, 4, ..., 8, 10 — gaps where models exist only in the Community table. Revert to per-table ranking: each table sorts and ranks against its own rows via sort_by_avg_rank. Same model in both tables can have different rank numbers (e.g., uAI-NEXUS-MedVLM-1.0d-4B-SFT at official-4 vs community-5 because the community table has more competitors). Both tables now display sequential ranks 1, 2, 3, ... with no gaps. The dominance tiebreaker added in commit 3a73928 (count of #1-metric wins) ensures top models like 1.0b-4B-RL stay at rank 1 across both tables when they win the most metrics outright. Removes three now-unused helpers: _read_leaderboard_json, _global_rank_lookup, _apply_global_rank. Verified against actual leaderboard JSONs: - Official (15 row…4ad37f6
-
Rank against the union of official + community to keep ranks consistent
MedGRPO Team +1: Claude Opus 4.7 (1M context)Same model was getting a different rank in the Official vs Community views because each view ran sort_by_avg_rank against its own competitor set (15 vs 35 models). With genuinely different avg-ranks per view, no amount of tiebreaking could make the displayed rank match. Fix: compute a single global ranking against the union of all submissions (official ∪ community), then map each view's rows to their global rank number. The same model now shows the same rank in either table; the Official table just skips rank numbers for rows that aren't in the official subset (e.g., rank 7 → 9 if rank 8 isn't promoted). Update the About-tab explanation to describe the global ranking. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>03adf33 -
Fix non-deterministic rank ties; correct metric descriptions
MedGRPO Team +1: Claude Opus 4.7 (1M context)Same model could rank differently in the Official vs Community tables because sort_by_avg_rank used mergesort's stability over input-row order as the implicit tiebreaker — and the two leaderboard JSONs have different insertion orders. Add 4 explicit tiebreakers: 1. Count of metrics where the model is ranked #1 (dominance) 2. Sum of per-metric ranks 3. Sum of normalized raw scores 4. Model name alphabetical (final fallback) Now a model with the same scores against the same competitor set gets the same rank regardless of which table loads it. Also fix the About-tab metric descriptions: replace mAP@0.5 (wrong metric for TAL) with the actual table column names (TAG_mIoU@0.3, TAG_mIoU@0.5), and add the missing RC_llm row so all 10 metrics in the table are described. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>3a73928
-
update
MedGRPO Teame51187b
-
update
MedGRPO Team1a7ba72 -
update logs
MedGRPO Team68d9818 -
update
MedGRPO Team0950b39
-
update
MedGRPO Teamd6c2cec
-
Reassign rank column after sort so medal badges match new ordering
MedGRPO Team +1: Claude Opus 4.7 (1M context)sort_by_avg_rank reordered rows but left each row's stored 'rank' field stale, so the medal badges (🥇🥈🥉) were driven by the old CVS-sorted rank values while the row ordering followed the new avg-rank sort. The visible result: row 3 in the table was rendered with a 🥉 that belonged to a different model. Fix: assign rank = [1..N] on the freshly-sorted DataFrame before return. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>fdadbd6 -
Rank leaderboard by average rank across all 10 metrics
MedGRPO Team +1: Claude Opus 4.7 (1M context)Previously sorted by cvs_acc only, which is misleading when top models cluster within 0.004 on that single metric. The new sort computes each model's rank per metric (1 = best; ties share smaller rank; NaN → last), averages those ranks, and sorts ascending (lower avg rank = better). - Add sort_by_avg_rank helper - Replace all 7 sort_values('cvs_acc', ascending=False) call sites - Add "How Models Are Ranked" section to the About tab - Refresh stale CVS-sort comments Robust to metric-scale differences (accuracy 0–1 vs LLM-judge 1–5) and rewards models that are strong across tasks rather than exceptional on one. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>faf76da -
Update About-tab citation: CVPR format, clearer ask, aligned sample count
MedGRPO Team +1: Claude Opus 4.7 (1M context)- Citation: @article{su2024medgrpo,...year=2025} → @inproceedings{su2026medgrpo,...booktitle=CVPR,year=2026} (matches model card and dataset README) - Lead-in: explicit "If you use our model or benchmark, please cite" - Training samples: 21,060 → 51,505 (matches project page) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>a690dc6 -
add model links
MedGRPO Team0b84800 -
update
MedGRPO Team4c12bb6 -
fairness
MedGRPO Teamb18e2e0 -
Update HF model references to uAI-NEXUS-MedVLM-1.0a-7B-RL
MedGRPO Team +1: Claude Opus 4.7 (1M context)Model was renamed from Qwen2.5-VL-7B-MedGRPO. Updates the model link in the leaderboard header, the submission-form placeholder, and the initial data upload script. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>039cb09 -
update is Ours
MedGRPO Team20d5553 -
update colors
MedGRPO Teamacee790 -
update
MedGRPO Teama571a44 -
update
MedGRPO Team16d284c -
update
MedGRPO Team1db37ad
-
Add official leaderboard, merge About tab, restructure Submit flow
MedGRPO Team +1: Claude Opus 4.6 (1M context)- Add Official Leaderboard tab for verified models (separate from community) - Merge Tasks & Metrics and About into single About tab - Move LLM Judge to Submit Results tab as clear Step 1/Step 2 flow - Add Model link to header - Fix sticky table headers and test set statistics formatting - Clean up community submissions language Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>0cba76d -
update
MedGRPO Team8eca10c -
update
MedGRPO Teameda88c2 -
update colors
MedGRPO Team89b65cf -
update
MedGRPO Teamd46ac00 -
add official leaderboard
MedGRPO Team66e9598
-
update
MedGRPO Team77926fe -
add color
MedGRPO Team7016445 -
update
MedGRPO Team3a49669 -
update
MedGRPO Team3f9e991 -
update
MedGRPO Teamadc43a0 -
update
MedGRPO Team05d6c42
-
fix issues
MedGRPO Teame2b1040
-
Update evaluate_predictions.py
MedGRPO Team +1: Claude Opus 4.6 (1M context)Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>a9b5dcf
-
update
MedGRPO Teamaa5db53
-
Update GitHub and project page URLs
MedGRPO Team +1: Claude Opus 4.6Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>63df552
-
add leaderboard printout
MedGRPO Teamc92c494
-
update
MedGRPO Team9aa148b -
relax
MedGRPO Team0098a7d