Shisa V2 Qwen2.5-32B: Bilingual Japanese/English Chat Model

Shisa V2 Qwen2.5-32B is a 32.8 billion parameter model from Shisa.AI, designed for high performance in both Japanese and English conversational tasks. It is part of the Shisa V2 family, which focuses on optimizing post-training with an expanded and refined synthetic-data driven approach, rather than tokenizer extension or continued pre-training. This model demonstrates significant improvements in Japanese language performance compared to its base model, Qwen2.5-32B-Instruct, across various benchmarks.

Key Capabilities & Differentiators

Bilingual Proficiency: Excels in Japanese language tasks while retaining robust English capabilities.
Optimized Post-Training: Achieves enhanced performance through a sophisticated synthetic-data driven approach, including extensive supervised fine-tuning (SFT) and DPO stages.
Extended Context Length: Features a 131,072 token context window, suitable for handling long conversations and complex documents.
Strong Benchmarking: Outperforms its base model in Japanese average scores (76.97 JA AVG vs. 66.79) and shows competitive results against other models in its class, particularly in JA MT Bench, Rakuda, and shisa-jp-rp-bench.
Custom Benchmarks: Development included new Japanese-specific evaluations like shisa-jp-ifeval (instruction-following), shisa-jp-rp-bench (role-play), and shisa-jp-tl-bench (translation proficiency).

Ideal Use Cases

Japanese-centric Applications: Suited for chatbots, customer service, and content generation requiring high-quality Japanese output.
Bilingual Communication: Effective for scenarios demanding seamless switching or translation between Japanese and English.
Long-Context Tasks: Benefits applications requiring extensive context understanding, such as document analysis or detailed conversational agents.
Role-Playing and Creative Writing: Demonstrates strong performance in Japanese role-play and creative tasks, with recommendations for higher temperatures (e.g., 1.0) for such uses.