Bringing AI Functions to the Llama Family

I’ve just released a dataset and 7B chat model to bring OpenAI-style AI functions to open source LLMs and the quasi-open Llama family of LLMs. I’ve set up a HuggingFace Space for people to try it out. I’ve been working on this project since Llama-2 was released roughly 2 weeks ago. I’ve been experimenting with utilizing AI functions to help structure reasoning and control flow in autonomous and quasi-autonomous agents over the last few months and I thought it would be interesting to see how much of that functionality I could bring to smaller models that were more open and commercially-flexible than OpenAI’s proprietary API.

Intially, I’m releasing a model that’s a fine-tuned version of the Llama-2 Chat model. I plan to iterate a bit more on the dataset and the 7B model before moving on to fine-tuning 13B & 70B chat models, as well as 7B, 13B, and 70B instruction-tuned models.

I will share more about my experience soon, but briefly, this was a really fun project. I was able to exercise a lot of the skills and knowledge I’ve been accumulating since late last year. It involved synthetic data generation, data augmentation, balancing multiple objectives in a single model (chat helpfulness, determining when and when not to call a function, and, when appropriate, accurately calling the correct function), evaluating model performance on a task for which there is no established eval, and fine-tuning with QLoRA.

Stay tuned for more details on my process, the code that I used to generate and augment the data, lessons learned, and my plans for continuing to evolve the dataset and trained models to improve accuracy and helpfulness and reduce hallucinations.