Function Calling Conversations in ShareGPT Format

86,864 chat examples designed for fine-tuning AI models on function calling, derived from the glaiveai/glaive-function-calling-v2 dataset. The dataset was created by hypervariance and last updated in December 2023. Each conversation includes a system prompt specifying 0, 1, or 2 available functions and instructions for the AI agent.

Use Cases

• Fine-tuning language models to generate correct JSON function calls based on the 'system' prompt specifying available functions.
• Training models to handle multi-turn conversations where a 'function_response' role provides results for a prior function call.
• Evaluating model performance on function invocation tasks with varying complexity (0, 1, or 2 functions provided in the context).
• Studying the structure of ShareGPT-formatted chats that integrate tool use as part of a dialogue flow.

Strengths

• 86,864 examples provide a substantial corpus for training.
• Explicitly includes a 'function_response' role to model tool output integration.
• System prompts define a controlled number of functions (0, 1, or 2) for structured learning.

Limitations

• Dataset size is fixed at 86,864 rows; no information on potential for growth or updates.
• Derived from a single source (glaiveai/glaive-function-calling-v2), which may limit topic diversity.
• Examples with invalid JSON as function calls were removed, which may reduce edge-case coverage.

Provenance

Source

Derived from the glaiveai/glaive-function-calling-v2 dataset.

Collection Method

Converted to ShareGPT chat format with modifications including adding a 'function_response' role and removing certain examples.

Freshness

Last updated on 2023-12-08.