Large language models have demonstrated remarkable emergent properties interacting directly with humans. These foundation models, many of which are instruction-tuned for chat are highly variable in their responses making it chal-
lenging to incorporate them for tasks without supervision.

OpenAI released a proprietary "Function Calling" model that addressed this issue by guaranteeing that the model inference would follow a structured JSON format. This deterministically structured output opens the door to LLM applications where human supervision is no longer required. For this research, I survey options on how to replicate the
"Function Calling" model on the open-source model Llama 2 chat.

Research summary

Development & results

canva

Poster

drive.google

Research Paper

Slide Presentation

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface

llama-7-chat-instruction-int4-fc-op_glaive-sft

Weights & Biases, developer tools for machine learning

wandb

Evaluation results

Replicate OpenAI's function-calling model. Contribute to CakeCrusher/gaitor-function-calling development by creating an account on GitHub.