Go Bruins V2.1: DPO Fine-tuned Language Model

Go Bruins V2.1 is a 7 billion parameter language model developed by Ryan Witzman, building upon the rwitz/go-bruins architecture. This iteration has been fine-tuned using Direct Preference Optimization (DPO) over 642 training steps, primarily on the athirdpath/DPO_Pairs-Roleplay-Alpaca-NSFW dataset. The model is designed to enhance text generation capabilities, language understanding, and sentiment analysis, with a focus on producing human-like text.

Key Capabilities

• Text Generation: Excels at generating diverse and human-like text.
• Language Understanding: Demonstrates proficiency in comprehending linguistic nuances.
• Sentiment Analysis: Capable of analyzing and identifying sentiment within text.
• DPO Fine-tuning: Leverages Direct Preference Optimization for improved performance over its predecessor.

Performance Metrics

Evaluations show competitive scores across various benchmarks:

• Average Score: 72.07
• Arc Challenge: 69.8
• Hella Swag: 87.05
• MMLU (5-shot): 64.75
• Truthful Q&A: 59.7
• Winogrande: 81.45
• GSM8k: 69.67

Usage Considerations

While versatile, users should be aware that due to its training data, the model may output NSFW or illegal content. It is not recommended for professional advice, crisis situations, or illegal activities. The model is licensed under MIT.