This research introduces a massive dataset with over 1 million instruction-API call pairs to dramatically improve how language models generate code.

• Fine-tuned models can outperform GPT-3.5 and GPT-4 when generating code for new API calls
• A fine-tuned CodeLlama-13B model demonstrates superior API call generation capabilities
• Results span multiple programming languages, making this advancement broadly applicable
• Enables more accurate automatic code generation for developers working with unfamiliar APIs

For engineering teams, this breakthrough means more reliable code assistants that can accurately suggest API implementations, reducing development time and errors when integrating external libraries.

API Pack: A Massive Multi-Programming Language Dataset for API Call Generation