Goliath 120B

Goliath 120B, an alpindale model created by merging two fine-tuned Llama 70B models (Xwin and Euryale), demonstrates exceptional performance in terms of speed and cost-efficiency. It consistently ranks among the fastest models and offers among the most competitive pricing across all evaluated benchmarks. However, its accuracy across various benchmarks presents a mixed picture. In Ethics and Mathematics, Goliath 120B shows significant weaknesses, achieving only 44.0% and 3.0% accuracy respectively, placing it in the lower percentiles for these categories. Similarly, its performance in Coding is modest at 22.0% accuracy. A notable weakness is observed in Instruction Following, where the model recorded 0.0% accuracy, indicating a substantial challenge in processing and executing complex directives. Conversely, the model exhibits a relative strength in Email Classification, achieving 95.0% accuracy, which, while not top-tier, is a respectable performance. The model's reliability is not explicitly provided in the rankings, preventing a specific comment on this aspect. Overall, Goliath 120B stands out for its operational efficiency but requires significant improvement in its ethical reasoning, mathematical capabilities, and instruction following.