拉風輸入法（Llavon IME）LLaMA 250M
This is a small decoder-only LLaMA-style language model trained for Traditional Chinese Bopomofo IME conversion.
The model predicts Traditional Chinese characters from Bopomofo input and is intended to be used as a local Windows IME inference model.
Project source code and IME implementations for Windows, macOS, and Linux: 拉風輸入法（Llavon IME）on GitHub.
GGUF builds are published separately in tony65535/llavon-ime-llama-250m-GGUF.
Model
- Architecture: LLaMA decoder-only Transformer
- Parameters: approximately 247.7M
- Context length: 384
- Hidden size: 1024
- Layers: 20
- Attention heads: 16
- RoPE theta: 10000.0
- Vocabulary: custom IME vocabulary with Bopomofo tokens and Traditional Chinese character tokens
Intended Use
This model is designed for local Bopomofo-to-Traditional-Chinese IME inference.
Typical inference format:
<BOS> <context-char-0> <context-char-1> ... <Bopomofo-0> <Bopomofo-1> ... <SEP> <char-0> <char-1> ...
The context characters before the Bopomofo tokens are optional already-entered Traditional Chinese text. During IME inference, logits should be masked to the candidate character set for the current Bopomofo syllable.
License
The model weights are released under CC-BY-NC-4.0.
The model weights were independently trained by this project. We do not take the position that model weights are necessarily reproductions, adaptations, or derivative works of the training datasets.
However, the training process included third-party datasets marked as CC-BY-NC-4.0. As a conservative approach to the original data licenses, source-rights uncertainty, and potential copyright risk, we release the model weights under CC-BY-NC-4.0 for non-commercial use only.
This licensing choice should not be interpreted as an admission that the model weights are necessarily governed by any particular training dataset license. It is a conservative release strategy adopted for risk management and open publication.
The source code for the IME runtime may be released under a separate software license.
Training Data Notice
The model was trained on multiple Traditional Chinese corpora, including public web text, conversation data, dictionary-derived entries, and PTT-derived user-generated text.
Some source datasets have incomplete, conflicting, or non-commercial licensing metadata. In particular:
yentinglin/TaiwanChatis licensedcc-by-nc-4.0.zake7749/Gossiping-Chinese-Corpus, the PTT-derived source used by this project, was observed as Apache-2.0 on GitHub.- Some web-crawl-derived datasets may be subject to upstream website terms or original author rights.
Users must comply with the CC-BY-NC-4.0 non-commercial license for the released weights and remain responsible for complying with applicable law, platform terms, and third-party rights.
The released model weights do not contain the original training corpora and are not intended to reproduce or redistribute the underlying training texts. This project does not grant any rights to use, copy, or redistribute the original training datasets.
See DATA_SOURCES.md for the training data statement and source summary.
Files
model.safetensors: Hugging Face / safetensors weights.ime_vocab.json,tokenizer.json,tokenizer_config.json: tokenizer assets.config.json,generation_config.json,training_args.json: model and training metadata.DATA_SOURCES.md: training data source and license evidence summary.
Status
Initial private upload.