lightblue/suzume-llama-3-8B-multilingual

Suzume-Llama-3-8B-Multilingual: Enhanced Multilingual Capabilities

lightblue/suzume-llama-3-8B-multilingual is an 8 billion parameter language model, fine-tuned from Meta's Llama 3 8B Instruct. While Llama 3 demonstrates strong English performance, this Suzume variant significantly expands its multilingual conversational abilities.

Key Capabilities & Differentiators

• Multilingual Fine-tuning: Enhanced with nearly 90,000 multilingual conversations, allowing it to respond effectively in various languages, unlike the base Llama 3 which often defaults to English.
• Strong Multilingual Benchmarks: Achieves competitive MT-Bench scores across 6 languages (German, French, Japanese, Russian, Chinese, English), often outperforming or matching models like Nexusflow/Starling-LM-7B-beta.
• Minimal English Degradation: Maintains strong English performance, with only minimal degradation compared to the original Llama 3 8B Instruct, while vastly improving non-English interaction.
• Training Data: Trained on a diverse dataset including lightblue/tagengo-gpt4 (76k conversations), megagonlabs/instruction_ja (669 Japanese conversations), and openchat/openchat_sharegpt4_dataset (6k multilingual conversations).

Good For

• Applications requiring a Llama 3-based model with robust multilingual conversational support.
• Chatbots and interactive AI systems targeting non-English speaking users or requiring mixed-language interactions.
• Developers seeking a powerful 8B model that balances strong English performance with broad language coverage.