Overview

rinna/qwen2.5-bakeneko-32b-instruct is a 32.8 billion parameter instruction-tuned model developed by rinna, building upon the Qwen2.5 Bakeneko 32B base model. It is specifically designed for Japanese language tasks and adheres to the Qwen2.5 chat format.

Key Capabilities & Training

This model was enhanced through a multi-stage training process:

• Model Merging: Instruction-following capabilities were added by merging the base model with a Chat Vector derived from Qwen/Qwen2.5-32B-Instruct and Qwen/Qwen2.5-32B. This process involved subtracting and adding parameter vectors, excluding the embedding layer.
• SimPO Refinement: Simple Preference Optimization (SimPO) was applied using rinna's internal dataset to further refine its performance.

Performance

Benchmarking results highlight its strong performance in Japanese language evaluations:

• Achieves 79.62 on the Japanese LM Evaluation Harness.
• Scores 8.17 on Japanese MT-Bench (first turn) and 7.66 on Japanese MT-Bench (multi-turn), outperforming Qwen/Qwen2.5-32B-Instruct in these metrics.

Good For

• Applications requiring high-quality Japanese instruction-following.
• Chatbots and conversational AI systems targeting the Japanese market.
• Tasks benefiting from a model optimized for Japanese language understanding and generation.