Shisa V2 Qwen2.5-7B Overview

Shisa V2 Qwen2.5-7B is a 7.6 billion parameter, Qwen2.5-based bilingual Japanese and English (JA/EN) chat model developed by Shisa.AI. It features a substantial 131072 token context length. This model is part of the Shisa V2 family, which focuses on optimizing post-training with significantly expanded and refined synthetic data approaches to enhance Japanese language performance while retaining robust English capabilities.

Key Capabilities & Differentiators

• Bilingual Proficiency: Designed for strong performance in both Japanese and English, with a particular emphasis on Japanese output quality.
• Optimized Post-Training: Achieves performance gains through extensive post-training and synthetic data refinement, rather than tokenizer extension or costly continued pre-training.
• Superior Japanese Benchmarks: Outperforms its base model, Qwen2.5-7B-Instruct, across numerous Japanese evaluation metrics including JA Avg (71.06 vs 65.30), Shaberi Avg, JA MT Bench, Rakuda, and Tengu.
• Extensive Evaluation: Tested using a custom "multieval" harness with standard benchmarks like ELYZA Tasks 100, JA MT-Bench, Rakuda, Tengu Bench, llm-jp-eval, MixEval, LiveBench, IFEval, and EvalPlus.
• New Japanese Benchmarks: Development of novel evaluations such as shisa-jp-ifeval (instruction-following), shisa-jp-rp-bench (role-play), and shisa-jp-tl-bench (translation proficiency) to measure specific Japanese downstream tasks.

Good For

• Applications requiring high-quality Japanese language generation and understanding.
• Bilingual chat and conversational AI systems where strong Japanese performance is critical.
• Tasks involving Japanese role-play, instruction-following, and translation, as indicated by specialized benchmarks.
• Developers seeking an openly-licensed (Apache 2.0) model with a large context window for bilingual use cases.