rime-frost (白霜拼音): a rebuilt Rime dictionary and schema set for Pinyin input

The problem rime-frost was built to fix

The README is unusually direct about its origin. The original configuration and dictionaries come from rime-ice at commit af2480b, and the stated complaint is that in the rime-ice dictionaries the character frequencies and word frequencies are not quite right and there are too many dead entries. A link to rime-ice issue 902 is given as the discussion thread behind that judgement.

The author's response was first to subtract. The README says unhealthy vocabulary was removed, along with a large number of obscure words defined as entries whose frequency equals 1 and which the segmenter cannot split, and entries such as 的吧 and 的了 that are not really words. Character and word frequencies were then edited by hand in quantity. Only after that came the statistical pass over a corpus of 745,396,750 characters, with segmentation, recounting of character and word frequencies, and normalisation.

That ordering matters for who should care. This is not a new input method and not a new engine. It is a dictionary and schema distribution for Rime, and the intended reader is someone who already has Rime running and has opinions about the order candidates appear in. Both full Pinyin and double Pinyin are supported, and the README states that the accuracy, on the project's own evaluation, no longer trails commercial input methods with cloud Pinyin disabled.

How the dictionary, schemas and Lua filters fit together

The repository layout shows the shape of the thing. rime_frost.schema.yaml is the main schema, with rime_frost.dict.yaml as its dictionary. Double Pinyin variants are separate schema files: rime_frost_double_pinyin.schema.yaml plus named variants for 自然码, 小鹤, 微软, 搜狗 and 紫光. There are also rime_frost_aux.schema.yaml for auxiliary codes, rime_frost_t9.schema.yaml for nine-key layouts, rime_frost_wubi86.schema.yaml, and unrelated schemas carried along from the same lineage such as bopomofo, cangjie5, stroke and melt_eng. Dictionaries live in cn_dicts, cn_dicts_cell, cn_dicts_common, cn_dicts_wb and en_dicts.

The Lua directory is where the behaviour beyond plain lookup lives. The README points at lua/aux_lookup_filter.lua for the in-sentence auxiliary code feature, and lists a set of trigger commands: /fh for symbols with the full list in symbols_v.yaml, /a /e /u for toned vowels, rq sj xq dt ts for date and time, uU for radical-based character lookup, U for Unicode characters, R for capitalised amounts, N for the lunar calendar, and V for a calculator. zh-moqi.gram is the language model file shipped in the repository.

The grammar configuration is a patch rather than a default, which is a deliberate choice worth noting: the model is not switched on for you. The README shows the patch keys (grammar/language, non_collocation_penalty, collocation_max_length, collocation_min_length, collocation_penalty, translator/contextual_suggestions, translator/max_homophones, translator/max_homographs) and states that if you would rather not edit rime_frost.custom.yaml you can download a model and replace the bundled zh-moqi. Model sources listed are octagram-data, rime-build-grammar and RIME-LMDG.

Installing rime-frost and typing your first sentence

The README gives two installation routes. The manual one is to download the repository as a ZIP, or the source-code.zip from the latest release, and unpack it into the Rime user directory for your platform. The paths differ per front end: on Windows, %APPDATA%\Moqi\Rime for the bundled 墨奇输入法 and %APPDATA%\Rime for 小狼毫; on macOS, ~/Library/Rime for 鼠须管 and ~/.local/share/fcitx5/rime for fcitx5-Mac; on Linux, ~/.local/share/fcitx5/rime for fcitx5-rime, ~/.var/app/org.fcitx.Fcitx5/data/fcitx5/rime for the fcitx5 flatpak build, and ~/.config/ibus/rime for ibus-rime. The README notes that %APPDATA% resolves to C:\Users\<USER>\AppData\Roaming.

The git route is shorter if you are comfortable with a terminal. The README says to cd to the parent of the configuration directory and clone into it, using Rime as the target folder name. On Windows in PowerShell that is:

cd "$env:APPDATA"
git clone --depth 1 https://github.com/gaboolic/rime-frost Rime

On macOS with 鼠须管 the parent is ~/Library:

cd ~/Library
git clone --depth 1 https://github.com/gaboolic/rime-frost Rime

After the files are in place you select 白霜拼音 in the input method settings and redeploy. The README states that after copying you should choose the schema in the settings and redeploy, which is the step that makes Rime read the new dictionaries. If you use 东风破 instead, the README lists recipes under others/recipes/ and gives the full install command as:

bash rime-install gaboolic/rime-frost:others/recipes/full

A first real use: the default schema is full Pinyin, so typing a normal Pinyin string produces candidates ranked by the rebuilt frequencies. The auxiliary code layer is optional and the README says it is opened with the backtick key and does not interfere with ordinary typing. The trigger commands listed above are typed the same way, so /fh opens the symbol set and rq inserts the date. Nothing in the README requires a model to be present for the base experience; the grammar patch is a separate, optional step.

Where rime-frost is the wrong choice

The most obvious boundary is that rime-frost is not an input method. It is a set of schema, dictionary and Lua files that a Rime front end consumes. The README's own framing is that the project mainly maintains dictionaries and word frequencies, and that a separate standalone Windows input method, 墨奇输入法, bundles it as the default schema. If you want something you install and start typing in without ever seeing a user directory, this is the wrong layer.

The second boundary is the update model. The README's update instructions are git pull inside the Rime folder, or downloading a ZIP again. It does not document rollback, and it does not describe how a merge would behave if you have local edits in the same files. Anyone who has customised rime_frost.custom.yaml, which the README itself suggests for the grammar patch, is editing a file inside the repository that a pull may touch. Keeping your customisations in a separate patch file, or keeping the user directory under version control so you can see what changed, is the practical mitigation, but the README does not spell this out.

Third, the frequency work is a claim backed by the project's own evaluation harness, rime-schema-compare, with a report generated 2026-04-25. The README states that rime-frost came out best both without a model and with the same model as the alternatives, and that it beats commercial input methods with cloud Pinyin off. That is a self-published benchmark with open, reproducible code, which is better than an unverifiable claim, but it is still one project's evaluation on one set of sentences. If your typing is dominated by a specialist vocabulary, the corpus statistics will not know your field any better than the next dictionary.

rime-frost against rime-ice, and against bundled input methods

The direct alternative is rime-ice, and the difference is not the engine or the schema syntax, since rime-frost's configuration descends from rime-ice commit af2480b. The difference is the dictionary and the frequency tables inside it. rime-ice keeps a broad set of entries with the frequencies it has; rime-frost deletes entries the author considers dead or non-words, hand-adjusts frequencies, and then re-derives them from a 745,396,750-character corpus. If rime-ice's candidate order works for you, switching buys you nothing except a different set of gaps. If you have ever wondered why a rare word outranks an obvious one, that is precisely the failure rime-frost was built to address.

The other alternative is the bundled route. 墨奇输入法 for Windows and Android ships rime-frost as the default schema, and 仓输入法 on iOS has it built in according to the README, with updates available through its scheme settings. That removes the file management entirely, at the cost of adopting another application and its release cadence rather than tracking the dictionary repository directly. The Android nightly build is distributed as a ZIP at a fixed releases/download/nightly URL, with the README noting a latest-release URL to fall back on if the nightly misbehaves, which tells you the author expects the nightly to occasionally break.

A third option in the same family is 薄荷拼音 and other schemes listed in the README's links, which consume the rime-frost dictionary without being rime-frost. That is worth knowing because it means the dictionary can be evaluated independently of the schema choices here.

Maintenance, licensing and what an upgrade costs you

The repository is not archived, and the last push was on 2026-09-23, so the dictionary is being touched. The release list is more informative than the push date: a nightly tag dated 2026-04-19, version 1.0.4 on 2026-05-08 described as adding support for the Android build of 墨奇输入法, and 1.0.3 on 2026-04-27 described as adding multi-sentence handling and auxiliary codes anywhere in a sentence. Two of those three releases are about the input method integration rather than the dictionary itself, which is consistent with a project whose centre of gravity has partly moved to the bundled application.

The practical upgrade cost is low but not zero. If you installed by git, updating is git pull in the Rime folder. If you installed from a ZIP, you unpack over the top. Either way the schema files, the cn_dicts directories and the Lua filters are replaced wholesale, and Rime must be redeployed for the change to take effect. The README does not describe a migration path for user dictionaries built up by typing, so treat your learned entries as the thing to protect before you overwrite anything.

Licensing: the repository is GPL-3.0. That is a copyleft licence, and it applies to a distribution that bundles configuration files, Lua code and dictionary data. Redistributing rime-frost inside another product, which is effectively what a bundled input method does, carries obligations under that licence. The README also credits the upstream sources it builds on, including rime-ice for the initial dictionary and most of the configuration, jieba, pypinyin, the MNBVC corpus, and kenlm. What that means for your own redistribution is a question for a lawyer, not for this article.

Editorial conclusion

Adopt rime-frost if you already run a Rime front end such as 小狼毫, 鼠须管, fcitx5-rime or 仓输入法 and the candidate order you get from your current dictionary bothers you; the switch is a folder replacement plus a redeploy, and the auxiliary-code layer is opt-in behind the backtick key. Do not adopt it if you want a packaged input method rather than a Rime configuration, or if you are not willing to keep the Rime user directory under version control, because the README's update path is git pull or a fresh ZIP overwrite and it does not document a rollback procedure. Before committing, read the rime-schema-compare report the README links to and decide whether its evaluation conditions match how you actually type.

Frequently asked questions

Official sources

1. gaboolic/rime-frost on GitHub
2. License: GPL-3.0
3. Project website
4. README
5. Releases