shisa-ai/shisa-v2.1-lfm2-1.2b
shisa-ai/shisa-v2.1-lfm2-1.2b is a 1.2 billion parameter bilingual Japanese and English chat model developed by Shisa.AI, featuring a 32K token context length. It is part of the Shisa V2.1 family, which focuses on enhanced Japanese language performance while maintaining strong English capabilities. This model is specifically optimized for local and edge-based use cases, offering class-leading Japanese-language performance at its size and significantly reduced cross-lingual token leakage.
Loading preview...
Overview
Shisa V2.1 is an update to the Shisa V2 family of bilingual Japanese and English (JA/EN) general-purpose chat models developed by Shisa.AI. This specific model, shisa-v2.1-lfm2-1.2b, is a 1.2 billion parameter variant with a 32K context length, designed for local and edge-based deployments. The V2.1 series incorporates an updated dataset, refined SFT and DPO recipes, and new additions for improved instruction following, translation, and politeness, particularly in Japanese.
Key Capabilities
- Bilingual Proficiency: Excels in Japanese language tasks while retaining robust English capabilities.
- Optimized for Edge Devices: Its 1.2B parameter size makes it suitable for local and edge-based use cases.
- Reduced Cross-Lingual Token Leakage (CLTL): Achieves a 13.5x improvement in reducing CLTL compared to its base model, ensuring more accurate and contextually appropriate Japanese output by minimizing the inclusion of non-Japanese tokens.
- Enhanced Instruction Following: Benefits from dataset and training improvements aimed at better instruction adherence and generation quality.
Good For
- Japanese Language Applications: Ideal for applications requiring high-quality Japanese text generation, chat, and translation.
- Resource-Constrained Environments: Suitable for deployment on devices with limited computational resources due to its compact size.
- Bilingual Chatbots and Assistants: Can power conversational AI systems that need to operate effectively in both Japanese and English, with a focus on Japanese accuracy.
- Research into Bilingual Model Behavior: Provides a strong baseline for studying and mitigating issues like cross-lingual token leakage in bilingual models.