Download VALIDATION.md from fffiloni/hf-picks: direct link, hf CLI and curl.
- Browser
- Download file 20.8 kB
-
https://huggingface.co/spaces/fffiloni/hf-picks/resolve/main/VALIDATION.md
- Command line
-
hf download hf://spaces/fffiloni/hf-picks/VALIDATION.md
-
curl -L -o VALIDATION.md https://huggingface.co/spaces/fffiloni/hf-picks/resolve/main/VALIDATION.md
Validation β v3.4.3
Consistent search controls
- All five content types share the same search-control styling and layout: a 44 px native select below the search field, then the result label and count. Spaces, models, datasets and blogs use the full row for their sort control; Papers retain their two-column period/sort row. Mobile text remains 16 px.
- Pasting a link hides the entire filter row without reserving an empty grid row. Blog discovery uses the same Newest first label and a concise scope note. Search options and backend ranking are unchanged.
- Existing Python/Node regression suites, frontend syntax and diff checks are used for this presentation update. No model inference or X interaction is required.
Validation β v3.4.2
Paper filter presentation
- Period and sort controls now form a compact row immediately below search, followed by the calendar scope and result count. Native selects keep keyboard and touch behavior, with a single consistent CSS chevron and 44 px targets. Mobile select text is 16 px to avoid iOS focus zoom.
- Removed the permanent instructional paragraph. Vote ranking retains a short visible scope note; full cumulative-vote and Daily Papers date semantics are available in About & data. The default paper sort reads Newest first.
- 54 Python and 32 Node regression tests pass; frontend syntax and diff whitespace checks pass. Search requests, ranking, period pagination, authentication, saved drafts and exported images are unchanged.
Validation β v3.4.1
Paper discovery periods β November 18, 2026
- 54 Python and 32 Node tests pass. New tests cover UTC calendar boundaries, complete-period pagination, winners beyond page one, deduplication, full-abstract and author matching, relevance, cumulative-vote ordering, cache reuse, upstream failures and invalid filters. Global keyword search remains unrestricted by date.
- Live public HF Daily Papers returned 34 papers today, 116 this week, 489 this month and 1,412 last month. All four period rankings returned 20 results in descending cumulative-upvote order. A subsequent monthly keyword search reused cached pages. Counts are a snapshot and can change.
- Longer periods fetch pages in small concurrent batches, sharing the existing bounded cache and request limit. A failed or incomplete period produces an explicit error instead of a partial top 20. A period has a maximum of 2,000 API candidates.
- Original paper publication dates, shelf rendering, draft storage and OAuth are unchanged. Existing authentication tests remain simulated. No GPU inference or X posting is involved.
Validation β v3.4.0
Automated and actual renderer
- 46 Python + 32 Node regression tests pass. This update adds shared primary/label/community baseline assertions for all five kinds, zero versus unavailable metadata, bounded community formatting, default text-only categories, and preservation of saved icon preferences. Existing authentication tests remain passing and simulated.
- A local Skia Canvas check using the bundled Source Sans 3 fonts rendered 36 native JPEGs: 12 normal/edge cases in each of Slim, Wide and Wider. Covers the five kinds, a long paper title, CPU Upgrade, RTX PRO 6000, missing hardware/date/community, zero counts and large counts. All detail strings stayed within their inset width without ellipsis or horizontal compression. Contact sheets were inspected at 720 px spine height. This is a controlled renderer check, not an X screenshot.
Published Chrome app β November 18, 2026
- Exercised all 12 width Γ Details Γ Rank combinations. The actual Download button produced 12 ZIPs with 72 valid JPEGs. Independently verified CRCs, six ordered image entries per archive, native 240 / 320 / 640 Γ 2048 dimensions, and manifest widths.
- Actual JPEGs range from 42,612 to 259,119 bytes. The official X image help accepts JPEG and lists a 5 MB photo ceiling. The app's files are below that ceiling; no claim is made about final X gallery layout, recompression, account-specific attachment limits or native share-sheet behavior. No post was published to X.
- Pixel masks of the community/detail blocks confirm the 128 px downward shift with Rank hidden in all 18 width/item comparisons (minimum mask overlap 97.8%, allowing JPEG compression). The rank separator disappears. Details hides both metadata slots and Wider context without leaving their reserved block.
- Category icons toggle independently. New shelves default to text-only category labels. Existing saved icon choices are retained.
- Responsive layout checked at 454 and 340 CSS px through browser zoom. No horizontal page overflow; all style labels have 44 px targets, and Width wraps cleanly at the smaller size. Zoom restored to 100%. This is not a physical-phone test.
- Restored and re-downloaded the original 33-item draft: title, options and every item match the saved original.
docs/interface.jpgshows the final Wide app;docs/slim-details.jpgshows Slim. Live runtime verified RUNNING on current/requested CPU-upgrade and health reports 3.4.0.
No GPU inference, private Hub content or X account interaction was used. OAuth behavior was not changed in this release; the v3.3.3 real-account completion limitation below still applies.
Validation β v3.3.3
Automated, local
46 Python and 30 Node tests pass, plus independent ZIP CRC/order/UTF-8 checks. Tests cover the direct OAuth redirect and cross-site rejection, state/PKCE/browser binding, session isolation and revocation, null opener references, same-tab return, remembered sessions, bounded title fitting, consistent insets, icon alignment, rank separators, one-line hardware/dates, width and metadata snapshot roundtrips, and overview size bounds. Provider responses and authenticated sessions in these tests are simulated.
Published app, November 18, 2026
- Three actual UI ZIP downloads contain 18 valid JPEGs: six each at 240 / 320 / 640 Γ 2048. Hardware and dates remain in the details block on one line; Slim puts their icon above the full-width value. Long paper titles and Slim side insets were visually inspected in the exported raster images.
- Earlier in this iteration, all 12 width Γ Details Γ Rank combinations were exercised and 72 v3.3.1 JPEGs verified. Rank/details spacing, arrow upvote icons and separators passed raster checks. Final v3.3.3 reran exports in all three widths and toggled Details/Rank; it does not claim a new 12-combination raster run.
- Final 20-pick Wider overview downloaded successfully at 8192 Γ 1459, below 12 megapixels; native individual-spine dimensions are unaffected. The six-pick Wider overview previously verified at 4064 Γ 2360. Shared-link roundtrip retained Wider and new public metadata.
- HF-hosted iframe: one click reaches the HF login page directly, with no intermediate app page; the original editor remains in its pending state. The browser returned a null popup reference although the tab opened, reproducing the opener isolation that broke the earlier handshake. This case is now explicitly supported and regression-tested.
- Direct app: one click navigates the SAME tab to HF login. Browser Back returns to the six-item editor with its draft intact; cancellation was checked through the account dialog. This browser had no signed-in HF account, so real consent completion, automatic return after consent and reuse of a real authenticated session remain unverified. Their application behavior is covered by simulated tests; no extra HF permission was requested.
- The original 33-item browser draft was restored and downloaded again. Title, options and every item match the saved original exactly. Screenshots show the final Wider interface (
docs/interface.jpg) and consistent Slim details (docs/slim-details.jpg). - Responsive checks earlier in this iteration at about 454 and 340 CSS px using browser zoom showed no page horizontal overflow; the shelf owns its horizontal scrolling. No new physical phone, touch-share or X-post test was performed.
No GPU inference was performed. Deployment uses the existing CPU-upgrade Space and short CPU upload Jobs. Server restarts still invalidate in-memory app sessions; public drafts persist separately.
Validation β v3.1.0
Weekly selection and sign-in return
- 39 Python and 11 Node tests pass. Six weekly-selection regressions cover the real reduced Hub HTML shape, editorial order, exclusion of trending results, ten-item limits, deduplication/backfilling, dated rollovers, empty rosters, changed contracts and anonymous access.
- Live adapter returned the eight official picks displayed for 16 Nov 2026, in matching order, with no warnings. It uses the current displayed roster rather than the historical featured API, which can include replaced entries.
- Fixed the sign-in tab that previously stayed open after success. Simulated frontend tests execute the actual callback script with no opener, verify successful handoff closes the temporary tab and focuses the original editor, keep errors visible, and verify cancellation closes the tab. OAuth transport/security tests remain passing. These are simulated tests, not a new real-account login.
- Deployed 3.1.0 on the existing CPU Upgrade Space. Chromium loaded eight weekly picks anonymously with the correct published date and source link; deselection changed Add 8 to Add 7, and switching to the demo returned five picks and hid the weekly source link. The existing 33-item draft stayed intact.
- The four-source selector was checked at 337 CSS pixels using browser zoom: document and dialog had no horizontal overflow and Add remained visible. This is not a physical-phone test.
- Chromium reached the real HF login page. Without entering credentials, the temporary tab was deliberately navigated to the success completion page to simulate only the final UI return. That tab closed itself, leaving the original editor tab and its 33 picks. The pending uncompleted OAuth attempt was then cancelled in the editor. This verifies browser tab closure, not a completed real-account login.
- No application-origin console errors were recorded. The updated screenshot in docs/interface.jpg shows the deployed weekly selector.
Earlier release verification
Validation β v3.0.0
Local verification β November 18, 2026
- 33 Python tests and 8 Node tests passed. New coverage includes merged chronology, ten-item example limits, deduplication/backfilling, creation versus modification dates, article authorship, explicit partial-source errors, paginated empty pages, unsafe references and credential exclusion from portable data.
- OAuth transport is simulated with HTTPX MockTransport. Tests cover exact scope/redirect/PKCE, secure callback cookies, browser/state binding, cross-origin start rejection, handoff verifier validation, callback replay, cancellation, expiry/capacity, provider errors, invalid identities, revocation and two independent accounts.
- Live public Hub adapters returned ten mixed favorites and ten latest creations for fffiloni, with no source warnings. The demo returned five items in editorial order.
- No GPU inference or social-network posting was run.
V3 deployment and Chromium verification
- Public Docker Space reports version 3.0.0, RUNNING, current/requested cpu-upgrade. HF automatically provisioned OAuth;
/api/auth/mereports it enabled. Only short CPU Jobs were used to publish files. - The V2 browser draft retained all 33 items, its title, settings and 13 reserve items after reload. Loading a preview did not change the draft.
- The demo preview returned five picks; deselecting one changed the add action to four. A real public-profile preview returned ten merged favorites with Liked/Upvoted dates, and ten creations with Created dates.
- Added nine of the ten favorites to the existing draft. Downloaded JSON independently confirmed all 33 original objects stayed unchanged, title/settings were preserved, nine unique items were appended, and account credentials were absent. Reopening the preview disabled the nine duplicates and offered only the remaining item.
- An unknown public profile produced an explicit error with Add disabled. Entering a valid profile recovered. The chosen public profile is visibly distinguished from a signed-in account; its form collapses after submission.
- Exported the ten live favorites through the real browser UI. Python independently verified ZIP CRCs, ten 240 Γ 2048 JPEGs, and a ten-item manifest spanning Spaces, models, a Paper and a blog post. Restored the original 33-item draft through the actual file chooser afterward.
- Responsive selector checked at 449 and 337 document CSS-pixel widths using browser zoom: no horizontal document/dialog overflow; the list scrolls, keyboard focus reaches the last checkbox, and the sticky Add control remains accessible. Zoom restored to 100%. This is not a physical-phone test.
- The live sign-in button opened the app's handshake tab and then the official HF login page with the registered client, correct app callback, profile-only scope, state and S256 PKCE. No HF browser account was connected during verification: the successful provider exchange/session handoff, logout and multi-account isolation were tested with simulated OAuth transport, not a real signed-in account. Real account consent/callback and the embedded iframe flow remain to be verified by signing in.
- Cancelling a second sign-in from the app closed the account sheet, restored the Sign in control and preserved the 33-item draft.
- No application-origin console errors were recorded.
docs/interface.jpgshows the deployed V3 selector with ten public-profile favorites.
Earlier release verification
Validation β v2.0.0
Local verification β November 18, 2026
- 16 Python regression tests and 7 Node tests pass, including both Collection formats, 33-item order, lazy metadata loading, partial errors, exact-match ranking, 20-image boundaries, reserve promotion, V1 migration, identity-based async updates, and explicit draft limits.
- ZIP integrity is checked with Pythonβs independent reader. Python compilation and JavaScript syntax checks pass.
- Live Hub metadata adapter checked with a 33-item public Collection and searches for all five content types.
V2 deployment and Chromium verification
- Public Docker app reports version 2.0.0 and runtime RUNNING on current/requested cpu-upgrade. Deployment used short CPU HF Jobs; no GPU inference or social-network publishing.
- The existing five-item V1 browser draft migrated without loss. The style controls start collapsed.
- Pasted a real 33-item Collection into the main search, reviewed its title/count, then imported it in author order. Observed the selected-metadata loading state followed by 20 enabled export picks and 13 interactive gray reserve spines.
- Promoted SAM2 Image Predictor from reserve into position 20. SALMONN Audio Questioning moved to position 21; total selection stayed at 33.
- Downloaded the actual 20-image ZIP through the UI. Python independently verified CRCs, all 20 JPEGs at 240 Γ 2048 pixels, numbered ordering and a 20-item manifest whose last pick is SAM2 Image Predictor. The reserve item is absent from the caption and exported manifest.
- Copied a snapshot link and opened its shared view (including same-tab fragment navigation): 20 colored spines, no reserve. Back to my draft restored all 33 items. Copied caption contains exactly 20 numbered picks.
- Save draft produced a 33-item JSON file. Cleared the test selection and restored that file through the real browser file chooser; title, first 20, reserve and ordering returned intact.
- A second Collection produced 42 unique items. Verified reserve pagination: 20 rows on page one, two on page two, with matching gray preview spines.
- Imported a deliberately unavailable public identifier, promoted it and observed the explicit metadata error with export disabled. Removing it restored the previous twentieth pick and enabled export.
- Verified exact owner/name search and the Most liked selector. Live adapter searches also covered models, datasets, papers and recent blog posts.
- Responsive layout checked at 449 and 337 CSS pixels using browser zoom, with no document overflow. The preview shortcut scrolled to the output, the 305-pixel dialog remained usable, and editing/style controls worked. Zoom restored to 100%. This is not a physical-phone test.
- Browser checks confirmed the final shortcut is hidden while output is visible and a reserve title typed during metadata loading is preserved.
- Overview download verified as a 5248 Γ 2360 PNG for 20 slim spines; the individual twentieth image is a 240 Γ 2048 JPEG.
- Additional V2 refinements: the preview shortcut hides when output is visible; reserve widths match the selected image width; reserve metadata loads when opened; async metadata does not overwrite a display title typed while loading; in-flight export edits invalidate prepared sharing files and stale progress.
Native phone share sheets, touch dragging and an actual X post remain untested. Reordering buttons, ZIP downloads and individual images provide the supported alternatives. No measured performance gain is claimed.
Previous release evidence
Validation β v1.0.2
Automated, local
- Python 3.12: 12 regression tests passed. Covers 30-day versus lifetime downloads, missing values, current versus requested hardware, declared/fallback palettes, paper wrappers, blog HTML contracts, URL validation, request coalescing/cache, upstream error handling, Collection order, partial failures, duplicate removal, ten-item limit and bundled font serving.
- Node 24: two tests passed for type-specific image facts and ZIP construction. Python's independent ZIP reader verifies CRCs, entry order, Unicode filenames and contents.
- Python compilation and JavaScript syntax checks passed.
- Production requirements pin the directly used packages and Starlette to the tested versions. A test-client deprecation warning was resolved by using HTTPX's documented ASGI transport; warnings are not suppressed.
Deployment and browser β November 18, 2026
- Public Docker Space deployed through short CPU HF Jobs. Hub runtime verified
RUNNINGwith both current and requested hardwarecpu-upgrade. No GPU work was run. /api/healthresponds successfully. Both font assets are included in the repository; the deployed regular font was downloaded and verified as a 15,696-byte WOFF2 binary, not an LFS pointer.- Live metadata resolution and discovery passed for all five content types. The example fetches actual public data, including Space-declared colors, ZeroGPU hardware, monthly downloads, publication dates and HF votes.
- Chromium: loaded the five-item example, added five more URLs in one import, observed the ten-item limit and disabled addition controls, reordered items, edited/reset titles and colors, changed width/details, refreshed metadata, and preserved the draft after an invalid-URL error.
- Browser persistence survived a reload. A copied shared snapshot opened with all ten items in read-only mode while the editing draft retained five items. Returning from the shared view restored the draft.
- Downloaded the actual ten-image ZIP through the interface. Independently decoded all ten JPEGs with Pillow: every image is 240 Γ 2048, ordered filenames agree with
selection.json, source links are included, and ZIP CRC validation passes. Inspected a short Space title and the long BERT paper title visually. - Individual image and overview downloads completed through the interface. The individual model image retains its rank from the selection.
- Responsive layout checked at 449 and 337 CSS-pixel widths using browser zoom, including dialog editing, ordering and appearance options. Document width equaled viewport width in both cases. Browser zoom was restored to 100%. This is a layout check, not physical-device emulation.
- Fixed an observed stale export-status label after editing the selection: edits now hide the previous export's progress and reset the native-share preparation label.
Still device-dependent
Native iOS/Android share sheets, browser-specific file-picker behavior, touch dragging, and an actual post on X have not been tested on physical devices/accounts. Reordering buttons are available alongside dragging. The application does not run model inference or post to social networks.
Blog discovery depends on the current public Hub HTML; its parser has fixture tests and was checked against live community and index pages. Future Hub markup changes may need an adapter update. Metadata counters can change after any check.