![]() It's fascinating to me how much confirmation bias comes in to play in topics like this. Here's the link to the paper, although I would recommend playing with the GitHub instead: For me the way of smartly creating synthetic dataset was a masterclass of itself. The paper also proposes an end to end automated evaluation for this technique. The last part is helpful to make the fine tuned models useful in the real world by telling it which APIs to call for particular tasks Finally a retrieval model is trained to fetch the most useful APIs to call during inference The open source models are fine-tuned on the instructions annotated above Depth First Search Strategy is much more useful here compared to Chain of Thought or ReAct techniques ![]() The authors create a new depth based technique that is more effective for API following which is key for making the step above ChatGPT is used to annotate and provide reasoning for the examples The low quality ones are filtered out by removing examples that don't work or are too slow The authors start by collecting a large set of API examples along with documentation It creates 16,000 synthetic examples and improves LLaMA-1 model to get performant at ChatGPT levels The paper aims to improve API following capabilities of open source models ![]() I think there is an underrated side to the paper as well of using Large Language Models effectively for data creation. ToolLLM paper has been incredibly popular for creating the largest API following dataset to fine tune models along with backbones. A masterclass in creating Synthetic Datasets using LLMs! □
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |