Spaces:
Running on Zero
Running on Zero
Commit History
2 days ago
-
Examples are five RAP-150 benchmark cases reproduced from their exact prompts
hanhaonan02 +1: Claude Fable 5.1- examples/<case>/: paper.pdf, our poster (PNG + self-contained HTML), prompt.json (exact test prompt), thumbnails/ and full-resolution assets/; meta.json carries the exact instruction, layout spec and canvas size - picking a card fills PDF, instruction, spec and canvas; Generate feeds the stored prompt verbatim when the PDF, instruction and spec are unchanged (matched by PDF hash), otherwise the regular parser runs - canvas accepts exact WxH sizes; replaced NExT-GPT / ConforMix with Deep Ensembles (ICML 2025) and Theory of Mind (ICLR 2023) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>9b03fbd
5 days ago
-
Author line: drop bare e-mail domains, attach combining marks to the following letter
hanhaonan02 +1: Claude Fable 5.1Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>f1b7085 -
Take the author line from the first page, not PDF metadata
hanhaonan02 +1: Claude Fable 5.1Without an authors line the model invented author names on a Space-generated poster. Lines between the title block and the abstract are used; e-mails and stray combining marks removed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>96e4fd5 -
One request per ZeroGPU step; keep small-caps titles intact
hanhaonan02 +1: Claude Fable 5.1- The ZeroGPU proxy token is bound to the HTTP request; parsing plus several 120 s steps in one request failed with 'Expired ZeroGPU proxy token'. Generate is now a chain: /analyze (CPU) then up to 10 chained /step events, one GPU step each; Continue re-enters the chain. - Title extraction joins touching spans and takes whole title lines, so ICLR small-caps titles no longer collapse to initials; falls back to PDF metadata when the result looks broken. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>033094c -
Accept any colour-picker format for the accent; replace examples with five stronger test-set posters
hanhaonan02 +1: Claude Fable 5.1- normalize_accent(): #RGB/#RRGGBB/rgb()/rgba()/hsl() -> #RRGGBB, default on failure (the picker could send rgba(...) and the old check raised 'Choose an accent color in #RRGGBB format') - Examples: MAESTRO (ICLR 2023), Gradient Leakage (ICLR 2022), Humanoid Motion (ICLR 2024), NExT-GPT (ICML 2024), ConforMix (NeurIPS 2025), each with PDF, PNG and self-contained HTML Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>77e6466
6 days ago
-
Publish-ready PDF to poster demo with validated BF16 example and safe continuation
Verified signature38c5f38 -
Add training-aligned PDF parser, safe rendering and resumable poster generation
Verified signaturec58c39d