Skip to content

History

Revisions

  • Document spoken-only attribution, labels-only review, request logs and hover help Mirrors docs/wiki in on22s/alexandria-audiobook2#1211 with wiki-style links.

    Michii committed Oct 13, 2026
    13a9ae6
  • Add Japanese user guides and wiki navigation

    @on22s on22s committed Oct 10, 2026
    d9ff9a9
  • Keep context instructions standalone while README changes are in review

    @on22s on22s committed Oct 10, 2026
    5274f86
  • Link beginner and Chinese installation routes from sidebar

    @on22s on22s committed Oct 10, 2026
    a7c4111
  • Qualify historical context measurement and link workload sizing

    @on22s on22s committed Oct 10, 2026
    325e2f3
  • Reconcile context sizing and qualify model-free test coverage

    @on22s on22s committed Oct 10, 2026
    a451b6c
  • Add beginner and Chinese routes; qualify quick-test coverage

    @on22s on22s committed Oct 10, 2026
    dcc6160
  • Clarify workload and VRAM context sizing; document optimizer profiles

    @on22s on22s committed Oct 10, 2026
    144468c
  • Publish the rewritten wiki from docs/wiki at e5d375e8 (#686, #687): clearer pages, corrected guidance, current results, Which model for your card

    pinokio committed Sep 28, 2026
    907cba6
  • Publish the Sep 22-23 results and correct the adapter guidance The published wiki had drifted behind docs/wiki/ in the repository. #657 said it added the Qwen3.8 UD-Q3_K_XL result to the wiki results table, but the wiki is a separate repository that the pull request never touched. Results: adds the Qwen3.8-27B UD-Q3_K_XL rights-clean paired row (95.2% -> 96.0%, +0.8 pp, +61/-45, p=0.145) with its Sun Also Rises training-overlap footnote, and the Nemotron 3 Ultra per-book API baseline table with its do-not-pool caveat. Purely additive: every existing row and every quant file size recorded in the last two wiki commits is untouched. The new row carries no file size because the Qwen3.8 UD-Q3_K_XL size is not recorded anywhere I could read, and the Muse UD-Q3_K_XL figure is a different model's file. Evaluation recipes: the LoRA arm is the adapter's declared candidate scale rather than a hardcoded 1.0, plus the rule that a pilot's non-default scale must be replicated over the full panel and every tested scale reported. Prompts and adapters: stops recommending the Gen 3 Muse recipe. It measured -2.0 on the four-book fixture; the only Muse adapter that has won is the earlier task4k-multin template-fixed run (+4.4), which was multi-entry and answer-only. Adds the rule the Gen 3 michel2_full set broke: train on the window shape the product serves, and check the window rather than the prompt variant name.

    pinokio committed Sep 24, 2026
    d569340
  • Record Muse Q4 file size

    pinokio committed Sep 23, 2026
    df4bdd2
  • Show quant sizes in results wiki

    pinokio committed Sep 23, 2026
    185829a
  • Add consolidated results wiki page

    pinokio committed Sep 23, 2026
    1a0cb51
  • Add persistent wiki navigation and footer

    pinokio committed Sep 20, 2026
    c14a807
  • Align wiki with original Alexandria documentation

    pinokio committed Sep 20, 2026
    301a277
  • Add project documentation wiki

    pinokio committed Sep 20, 2026
    763e9b1
  • Initial Home page

    @on22s on22s committed Sep 20, 2026
    bc0fcf1