Qwen3-Swallow-8B-RL-v0.2 Overview

The Qwen3-Swallow-8B-RL-v0.2 is an 8 billion parameter model from the Qwen3-Swallow family, developed by tokyotech-llm. This model is a bilingual Japanese-English LLM, refined through a multi-stage training process including Continual Pre-Training (CPT), Supervised Fine-Tuning (SFT), and Reinforcement Learning with Verifiable Rewards (RLVR). The development focused on enhancing Japanese language and Japanese-English translation capabilities.

Key Capabilities

• Bilingual Proficiency: Highly optimized for both Japanese and English language tasks.
• Retained STEM Performance: Strategic CPT and SFT pipelines prevented catastrophic forgetting in mathematics and coding, utilizing high-quality math and code datasets with reasoning traces.
• Enhanced Reasoning: Achieves reasoning performance on par with, and in some cases surpassing, the original Qwen3 models, particularly in math and coding.
• RL-Tuned: The 'RL' in its name signifies its refinement through Reinforcement Learning, further boosting its reasoning abilities.

Good For

• Applications requiring strong Japanese and English language understanding and generation.
• Tasks involving Japanese-English translation.
• Use cases demanding robust mathematical and coding reasoning capabilities.
• Developers seeking a model with enhanced reasoning performance in STEM fields, built upon the Qwen3 architecture.