lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top75

Overview

lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top75 is an 8 billion parameter multilingual language model developed by lightblue. It is a fine-tuned version of the lightblue/suzume-llama-3-8B-multilingual base model, utilizing the ORPO (Optimized Ranking Preference Optimization) training method. The model was trained using a specific subset of the lightblue/mitsu dataset, focusing on the top/bottom responses from the 75% most consistently ranked prompts.

Key Capabilities & Performance

This model shows significant improvements in multilingual performance, particularly in MT-Bench scores across various languages. For instance, it achieved the highest MT-Bench score in Chinese (7.77) among the evaluated models, and strong scores in German (7.64) and Russian (8.74). While it performs well in English (7.94), other ORPO variants or proprietary models might slightly surpass it in specific languages. The training involved a sequence length of 8192 tokens and a learning rate of 8e-6 over one epoch.

Important Considerations

• License: This model carries a non-commercial license due to the use of Command R and Command R+ models for generating its training data (lightblue/mitsu dataset).
• Recommended Alternative: For general use, lightblue recommends their lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half model, which demonstrated the highest overall performance in their tests.

Good for

• Multilingual conversational AI applications, especially in Chinese, German, and Russian.
• Research and experimentation with ORPO fine-tuning techniques on multilingual datasets.
• Developers seeking a Llama 3-based model with enhanced multilingual understanding and generation capabilities for non-commercial projects.