Invoker-13B: Function-Calling Llama-2 Model
Invoker-13B is a 13 billion parameter large language model developed by jeffrey-fong, built upon the Llama-2 architecture. Its primary distinction lies in its advanced capability to intelligently plan between executing function calls and generating direct conversational responses, mirroring the behavior of OpenAI's function-calling models.
Key Capabilities
- Intelligent Function Calling: The model can analyze user queries and a provided list of functions (in OpenAI's JSON format) to determine the most appropriate function to call, or if a direct response is needed.
- Response Summarization: After a function call, it can summarize the function's output to provide a coherent answer to the user.
- Context Length: Fine-tuned with a 4096 token sequence length, enabling it to handle moderately long interactions and function descriptions.
- Efficient Training: Utilizes QLoRA and DeepSpeed Zero Stage 2 for reduced computational resource requirements during training.
Training Data & Methodology
The model was trained on a diverse dataset to enhance both conversational fluency and function-calling proficiency:
- ToolBench (0830 updated): A large-scale, high-quality instruction tuning dataset specifically for general tool-use capability, with rigorous cleaning to ensure relevant function calls and summarized responses.
- ShareGPT-34K: A filtered dataset of high-quality multi-turn conversations.
- OASST1: A human-generated, human-annotated assistant-style conversation corpus, filtered for English conversations.
Usage Considerations
- Requires approximately 1x A100 40GB GPU for full float16 precision inference.
- Adheres to a specific prompt format that includes a list of available functions or
None if no functions are present.
Invoker-13B is particularly well-suited for applications requiring an LLM to interact with external tools or APIs in a structured and intelligent manner.