TinyAgent aims to enable complex reasoning and function calling capabilities in Small Language Models (SLMs) that can be deployed securely and privately at the edge. Traditional Large Language Models (LLMs) like GPT-4 and Gemini-1.5, while powerful, are often too large and resource-intensive for edge deployment, posing challenges in terms of privacy, connectivity, and latency. TinyAgent addresses these challenges by training specialized SLMs with high-quality, curated data, and focusing on function calling with
LLMCompiler. As a driving application, TinyAgent can interact with various MacOS applications, assisting users with day-to-day tasks such as composing emails, managing contacts, scheduling calendar events, and organizing Zoom meetings.
Please see our
Github for details on how to use TinyAgent models. TinyAgent models can be used programmatically or through our user interface.
Dataset:
We curated a
dataset of
40,000 real-life use cases. We use GPT-3.5-Turbo to generate real-world instructions. These are then used to obtain synthetic execution plans using GPT-4-Turbo. Please check out our
blog post for more details on our dataset.
Using the synthetic data generation process described above, we use parameter-efficient fine-tuning with LoRA to fine-tune the base models for 3 epochs. Please check out our
blog post for more details on our fine-tuning procedure.
When faced with challenging tasks, SLM agents require appropriate tools and in-context examples to guide them. If the model sees irrelevant examples, it can hallucinate. Likewise, if the model sees the descriptions of the tools that it doesn’t need, it usually gets confused, and these tools take up unnecessary prompt space. To tackle this, TinyAgent uses ToolRAG to retrieve the best tools and examples suited for a given query. This process has minimal latency and increases the accuracy of TinyAgent substantially. Please take a look at our
blog post and our
ToolRAG model for more details.
