Spaces:
Running
Running
Commit History
-
Trigger rebuild (build job stuck in queue) Running
Ziyang Wang +1: Claude Fable 5Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>b0003f2 -
Disable experimental SSR — fixes Spaces health-check launch timeout
Ziyang Wang +1: Claude Fable 5Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>43aff03
-
Restore anonymous Score tab; keep leaderboard curation manual
Ziyang Wang +1: Claude Opus 4.7Corrects the previous "viewer-only" commit. The intended design: - Anyone can upload a submission.json to the Space's Score tab and see per-split + Overall accuracy against the held-out answer key. - That score is returned to the submitter only — nothing is persisted on the Space, and nothing is posted to the public leaderboard. - To appear on the leaderboard, the submitter must email <ziyangw@cs.unc.edu>; the maintainer verifies and pushes the row via scripts/publish_submission.py. Concretely: - app.py: add "Score" tab wired to evaluator.score_submission; no HF login required; the result markdown explicitly states nothing is persisted and links out to the email flow. - ledger.py: restore ensure_private_annotations() so the Space can pull annotations_private.json at boot for scoring. Still no write path. - README.md: HF_TOKEN scope is now "Read on Ted412/EgoMemReason-Private" only — no write scope anywhere. "Accepted at COLM 2026" added. Co…b4db1ef -
Switch Space to viewer-only + v1.1 revision notice
Ziyang Wang +1: Claude Opus 4.7The Space is now a read-only viewer of Ted412/EgoMemReason-Leaderboard. Self-serve submissions have been removed after a third-party user exfiltrated the private answer key and posted a 100% run (2026-06-17). Every leaderboard entry now has to be personally scored by the maintainer against the held-out key, then pushed via scripts/publish_submission.py. - app.py: strip Submit / Manage / Review tabs; keep About + Leaderboard. - ledger.py: shrink to list_submissions() reading the public dataset with no HF token. - evaluator.py: round returned metrics to 1 decimal (drops one bit of info per split per invocation). - auth.py, tests/: deleted — obsolete under the viewer-only design. - scripts/publish_submission.py: new maintainer-only CLI. Validates a submission against the local annotations_private.json and uploads a leaderboard record. - scripts/reshuffle_options.py: new tooling to permute per-question letter mappings, invalidating any leaked answer key. - Revision notice for v1.1…c83ff60
-
add arXiv paper
Ziyang Wanglink + update BibTeX1b38b03
-
add project-page link (egomemreason.github.io)
Ziyang Wang7f6d4bc -
update GitHub URL to Ziyang412/EgoMemReason
Ziyang Wang9cf02ac -
also patch _json_schema_to_python_type for bool schemas
Ziyang Wang0385d22 -
make About the default landing tab
Ziyang Wang483161e -
monkey-patch gradio_client.utils.get_type to handle bool schemas
Ziyang Wang7b3d33e -
pin gradio==5.15.0 — 5.0.0 still has the gradio_client schema bug
Ziyang Wang9861d7d -
tighten leaderboard: drop date col, shorten headers, no theme, no wrap
Ziyang Wanged7ed80
-
pin huggingface_hub<1.0 (Gradio 5 OAuth still imports HfFolder)
Ziyang Wang0c51b7f -
upgrade to Gradio 5.x to fix client utils schema bug
Ziyang Wang4f21599 -
fix py3.13 audioop dep: pin python_version=3.12 + audioop-lts
Ziyang Wang60b3d29 -
fix Space frontmatter: drop deprecated hf_oauth_scopes
Ziyang Wang5cf1c5a -
add paper-baseline seed script
Ziyang Wang6375fde -
initial Space
Ziyang Wang1bf5b23