chihoonlee10/T3Q-ko-solar-dpo-v6.0

T3Q-ko-solar-dpo-v6.0: DPO Fine-tuned Korean Language Model

This model, developed by Chihoon Lee and T3Q, is a 10.7 billion parameter language model that builds upon its predecessor, T3Q-ko-solar-dpo-v5.0. The key differentiator for v6.0 is its fine-tuning using Direct Preference Optimization (DPO), a method designed to align model outputs more closely with human preferences.

Key Capabilities & Performance

The model is specifically designed for Korean language processing and has been evaluated on several Korean benchmarks, demonstrating its proficiency in various NLP tasks. Performance metrics include:

• kobest_boolq: Achieves an accuracy of 0.5028 and a macro_f1 of 0.3396, indicating its ability in boolean question answering.
• kobest_copa: Shows strong common sense reasoning with an accuracy of 0.8020 and a macro_f1 of 0.8018.
• kobest_hellaswag: Records an accuracy of 0.5340 and a normalized accuracy of 0.5720, reflecting its performance in situational common sense tasks.
• kobest_sentineg: Demonstrates an accuracy of 0.7985 and a macro_f1 of 0.7956 for sentiment analysis.

When to Use This Model

This model is particularly well-suited for applications requiring a robust and preference-aligned Korean language model. Its DPO fine-tuning suggests improved conversational quality and adherence to desired output styles. Consider using T3Q-ko-solar-dpo-v6.0 for:

• Korean-centric chatbots and conversational AI.
• Question answering systems in Korean.
• Applications requiring common sense reasoning in Korean contexts.
• Sentiment analysis and text generation tasks for the Korean language.