Dracarys-72B-Instruct Overview

Dracarys-72B-Instruct is a 72.7 billion parameter language model developed by Abacus.AI, building upon the Qwen2-72B-Instruct architecture. It is part of the "Smaug" series, specifically targeting enhancements in coding performance across various base models. This variant demonstrates notable improvements in code-related tasks compared to its foundational model.

Key Capabilities & Performance

• Enhanced Coding Performance: Dracarys-72B-Instruct shows superior performance on the LiveCodeBench evaluation suite.
  • Code Generation: Achieves 33.57, outperforming Qwen2-72B-Instruct's 32.92.
  • Code Execution: Scores 62.96, significantly higher than Qwen2-72B-Instruct's 58.95.
  • Test Output Prediction: Reaches 58.93, compared to 55.88 by the base model.
• LiveBench Evaluation: Demonstrates a higher Global Average (41.20 vs 40.15) and Coding Average (38.95 vs 32.38) on the July LiveBench update.

Use Cases

• Data Science Coding Assistant: Excels at generating Python code, particularly for tasks involving libraries like Pandas and NumPy.
• Code Generation and Completion: Suitable for applications requiring robust and accurate code output.
• Educational Tools: Can be used in environments where code problem-solving and explanation are needed.