Fetching metadata from the HF Docker repository...
-
UI: keep fullscreen on the working overlay (display) image; only the editable input image drops it
-
UI: drop the broken fullscreen image button (Gradio 6 buttons=[download]); keep download
-
README: condense the radiologist section to a short factual note; full quote moved to the blog
-
UI: match Clear button height to Run audit (add font-size + vertical padding)
-
README: link the Nemotron draft-parser model in Links (parity with the grounding model)
-
README: add achievement:offgrid (no external cloud APIs; models run in-Space), claim Off the Grid
-
README: order social links X first, LinkedIn second
-
README: add X/Twitter social post link alongside LinkedIn
-
README: add Tiny Titan + Best Demo badge tags; sync radiologist quote with published blog
-
Mobile: cap container to viewport so wide JSON/table scroll internally; width-driven overlay; fix ribbon dot wrap
-
Stack columns on mobile (<=700px) and add (N foci) hint for multi-box findings
-
ui: point header blog chip at the published blog post
-
ui: header chip is now BLOG POST (links to the published Field Notes), replacing README
-
docs: add LinkedIn social-post link to README
-
docs: add published Field Notes link; align radiologist quote with the blog; drop 'honest' wording
-
docs: reframe radiologist section (a radiologist's need, not 'agreed to try'); honest placeholder
-
docs: attribute cxr_example_4 (CC-BY-SA); four examples
-
ui: filled README chip + hover; footer BUILD SMALL HACKATHON (all-caps)
-
ui: README chip in header; 4th (urgent) example; caption
-
examples: add urgent mass example (cxr_example_4, CC-BY-SA James Heilman)
-
docs: examples now all carry drafts (clean/missed/over-called)
-
fix: silence Nemotron top_p generation-config warning; document Mamba fallback
-
ui: examples now show clean/missed/over-called; pass model URLs to header
-
ui: clickable header chips + footer model/hackathon links; evidence-card height aligns with draft card
-
ui: drop unsupported from image legend (never drawn as a box); keep urgent emphasis
-
docs: assemble_audit/run_audit docstrings reflect Nemotron-on-GPU (not CPU/llama.cpp)
-
docs: draft-parsing prompt targets Nemotron; grounding prompt targets MedGemma
-
docs: parser is model-agnostic; serving model is Nemotron (not MedGemma)
-
docs: Nemotron now on GPU via transformers; drop Llama Champion tag/claim (llama.cpp removed)
-
fix: re-enable Nemotron reasoning (load-bearing for parse accuracy); widen token budget
-
fix: read input_ids from BatchEncoding (apply_chat_template returns dict, not tensor)
-
Both models on GPU; header/footer name both; faster audit
-
Drop llama-cpp-python; pin transformers>=5.3 for native Nemotron
-
Header/footer: show both models (grounding + draft) prominently
-
Draft parser: run Nemotron on GPU via transformers (drop llama.cpp/CPU)
-
Add hackathon tags + write-up to Space README (REQ-06): track/badge tags, two-model architecture, honest eval, links, radiologist placeholder
-
Honest audit-time note (grounding GPU + Nemotron CPU, ~1 min)
-
Draft parser: use all CPU threads for faster llama.cpp prompt eval
-
Sync cxr_auditor: Nemotron draft parser + assemble_audit split
-
Wire Nemotron draft parser into the audit handler
-
Add llama-cpp-python (CPU) for the Nemotron draft parser
-
Light theme: upload dropzone follows the paper palette (input adaptive, output viewer stays near-black)
-
Wire head script (default-dark + theme toggle) into launch
-
Adaptive theme: Clinical Light + Reading Room with header toggle (var-based colors)
-
Findings table: drop the no_finding sentinel as a row (mutually exclusive with positives); show one No finding row only for a clear study
-
Layout: full-width Audit verdict / Findings / Raw JSON stack under the evidence panel
-
Fix inconsistent panel corner rounding: round .styler clip wrapper and image-block bottom corners to the panel radius token
-
Bump Gradio to 6.18.0; launch(theme/css) requires Gradio 6
-
Reading Room UI redesign; move theme/css from Blocks constructor to launch() per Gradio 6 migration
-
Deploy collision-aware overlay label placement (render.py); sync cxr_auditor package with local sources
