[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1l37qurz2t0rl":3},{"code":4,"message":5,"data":6,"timestamp":41},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":16,"summary_md":17,"topic":18,"github_topics":21,"license":37,"homepage":38,"updated_at":39,"history":40},"0c-biggo\u002F0c-ime",9,"免選字的注音輸入法：在本機用 Transformer 看前文選字。支援 Linux（Fcitx5／IBus）、macOS、Windows。","Go",141,15,26,5,[],"An on-device Zhuyin input method for Traditional Chinese that uses a small Transformer to pick characters from context so you can type whole sentences without selection.","If you type Traditional Chinese with Zhuyin, 0C (靈犀注音) is worth a look as a daily-driver experiment rather than a drop-in replacement you assume will work everywhere. It targets the routine break in typing where you finish a syllable string and then stop to pick the right homophone.\n\n## The problem\n\nThe usual workflow is to type a few syllables, wait for the candidate list, and press a number key, or go back and correct a character after the sentence reads wrong. This interruption repeats several times per sentence and gets worse for names and technical terms. 0C keeps a statistical word-graph decoder but adds a Transformer that rereads the preceding 64 characters and rescores the top 8 candidates, so the first candidate is right more often and you can keep typing.\n\n## What's different\n\nA unigram\u002Fbigram model plus Viterbi proposes a short N-best list, and a causal Transformer decoder (8 layers, width 512, about 32.6 million parameters, trained from scratch) rescores it using both the typed Zhuyin and the text already committed. A Tab hint appears only when the model is confident, and it draws on dictionaries and personal phrases rather than generated text. Everything runs on CPU; both models total about 250 MB, with reported P95 latency of roughly 7–16 ms.\n\nThe key trade-off is the [evaluation](https:\u002F\u002F0c.biggo.com\u002Ftechnology\u002F): a frozen test of 6,000 Taiwanese-news cases from 2026-09-24 to 26, half difficult homophone phrases, with readings synthesized from a dictionary and answers taken from source text. The Transformer variant reached 86.0–86.1% first-choice accuracy versus 76.9% for the statistical engine alone and 64.2%\u002F67.2% for 新酷音 0.11.0 and 小麥注音. That is a real signal, but an author-run benchmark with synthetic input, not human typing; the README says selection-free typing is a goal, not a guarantee. No network calls, no upload of typing content, and no text sent to the model in password fields, paused learning or alphanumeric mode. Automatic fallback keeps typing usable if the model fails.\n\n## Who should try it\n\nGood fit: a Zhuyin user on Linux (Fcitx5 or IBus), macOS or Windows who writes long Traditional Chinese sentences and will install a release, add personal vocabulary and try the optional punctuation. Poor fit: someone who needs a battle-tested IME with years of reports, or who must use a platform outside the four documented shells and will not build and test it.\n\n## Requirements and limitations\n\nBuilding needs Go 1.26 and Python 3.11 for tooling; each shell has its own toolchain and libraries, as listed in the [build notes](https:\u002F\u002Fgithub.com\u002F0c-biggo\u002F0c-ime\u002Fblob\u002Fmain\u002FREADME.md). `make assets` pulls about 275 MB of dictionaries and paired models from the matching release, or about 36 MB for dictionaries only. Personal learning is local, the license is AGPL-3.0-only with a commercial option, and punctuation-on-commit is Beta. Rare names and out-of-vocabulary words may still need manual selection, and dictionary and model versions must match the installed release.\n\n## HotRepo's take\n\nHotRepo first tracked the repository at 5 days old; it appeared at #11 on the go NEW-REPOSITORY board (not the overall Trending list) on 2026-10-25 with 141 total stars, and the sample spans only 2026-10-24 to 2026-10-25, so persistence cannot yet be judged. Before adopting it, verify the shell for your exact platform and the dictionary\u002Fmodel pairing for the release.",{"slug":19,"name":20},"productivity","Productivity & Utilities",[22,23,24,25,26,27,28,29,30,31,32,33,34,35,36],"bopomofo","chinese-input-method","fcitx5","go","ibus","ime","input-method","language-model","macos","onnx","taiwan","traditional-chinese","transformer","windows","zhuyin","AGPL-3.0","https:\u002F\u002F0c.biggo.com","2026-10-25T03:13:00Z",[],1792915219]