How Does Hammer 2.1 Improve Multi-Step Function Call Capabilities Without Specific Training Data?
Thank you to the Hammer development team for this exciting project. I have reviewed the Salesforce/xlam-function-calling-60k and MadeAgents/xlam-irrelevance-7.5k datasets, but I have not found any specific data related to multi-step function calls. I also processed the training data using the Python script train/data_processing.py provided in the repository, but it does not appear to generate multi-step-related examples either.
However, on the Hugging Face page for Hammer 2.1, it is mentioned that Hammer 2.1 has significantly improved its capabilities in multi-step function calls. I would like to ask, how was Hammer 2.1 able to achieve such improvements in multi-step function call abilities without access to multi-step-specific training data?
Thank you for your time, and I look forward to your response!
Thank you for your interest and inquiry! Hammer 2.1 improves upon Hammer 2.0 by incorporating a small amount of multi-step and multi-turn training data. However, this specific data has not been open-sourced yet. We are currently working on further optimizations, and we plan to release this portion of the data in the future. Stay tuned!
OK Thanks for your detail reply