Details on SFT Data Generation and Training Pipeline

[VoyageWang]

Hi @authors, thanks for the excellent work on AORCHESTRA! The framework design is very insightful, especially the four-tuple abstraction for dynamic agent creation.

I'm trying to reproduce the SFT results reported in the paper (Section 3.3 and Table 1), and I noticed that some implementation details regarding the supervised fine-tuning (SFT) process are not fully described in the current version. To ensure correct reproduction, I'd appreciate it if you could clarify the following:

1. Training Data Construction

• Source of trajectories: Are the training trajectories collected from the training-free version of AORCHESTRA, or from human demonstrations, or from another method?
• Data format: What is the exact format of the training samples? Is it:
  • Prompt: Full history s_t
  • Completion: Action a_t (i.e., the Delegate(Φ) with the 4-tuple configuration)?
  • Or a different format?
• Positive/negative sampling: The paper mentions behavior cloning from expert trajectories. How are "expert" actions selected? Are they only successful trajectories, or do you also include failure cases with corrected labels?

Are there plans to release:

• The training data generation scripts?
• The processed SFT dataset?
• The training configuration and launcher scripts?

[CurryxIaoHu]

@Aurorra1123 Please solve this issue when you are free